WEBVTT
1
00:00:10.349 --> 00:00:13.789
Hi, welcome to the data standard
and I'm your guest host for today.
2
00:00:13.869 --> 00:00:16.789
Give you in Ubon. On our
show today we have Mickey along, the
3
00:00:16.910 --> 00:00:20.350
CTEO and founder, at gain site. Today we're going to be speaking about
4
00:00:20.750 --> 00:00:24.460
using data for marketing animation. Welcome
to the show. Hi, thank you
5
00:00:24.539 --> 00:00:28.660
for having me. Where are you
located right now? I'm from South Bay.
6
00:00:28.940 --> 00:00:33.659
Moved from Tel Aviv, Israel,
eight years ago. It's part of
7
00:00:34.539 --> 00:00:40.289
building a local presence to a former
start up. Actually, I founded nice
8
00:00:40.450 --> 00:00:43.049
and tell us a little bit about
your background. I'd love to hear more.
9
00:00:43.170 --> 00:00:46.689
Awesome. So my Baker is kind
of from computer science and also did
10
00:00:46.770 --> 00:00:51.799
my MBA degree, but started coding
at age thirteen or less. Kind of
11
00:00:51.920 --> 00:00:57.520
like Geeky with regards to computer science
and building actually software and products. I've
12
00:00:57.640 --> 00:01:03.629
dealt couple of years with retaime ONA
lyrics in the financial markets and then also
13
00:01:04.469 --> 00:01:10.989
started the two companies. One was
about website, web and brand experience.
14
00:01:11.629 --> 00:01:15.989
That kind of applied the same elements
of taking decision in real time and delivering
15
00:01:17.030 --> 00:01:23.700
a better experience for customers, and
now recently co found gains at Px,
16
00:01:23.819 --> 00:01:26.819
the product experience platform. That is
the same. Is just dealing with a
17
00:01:26.900 --> 00:01:30.620
lot of data, taking the those
insights, driving results and just delivering better
18
00:01:30.659 --> 00:01:34.489
customer experience in the end. Yeah, it so it sounds like you've done
19
00:01:34.530 --> 00:01:38.689
some similar things over the years.
Do you have best practices that you would
20
00:01:38.769 --> 00:01:45.290
share that you think you've learned over
out of these different iterations of problems?
21
00:01:45.609 --> 00:01:49.480
It's always about about the data in
the end, it's and you want to
22
00:01:49.560 --> 00:01:53.640
make sure that you're driving outcome for
your customers. So, you know,
23
00:01:53.719 --> 00:01:57.040
one of the best practices is obviously
always focus all the on the problem and,
24
00:01:57.400 --> 00:02:01.150
as they always say, try to
build something that is en x better.
25
00:02:02.109 --> 00:02:06.510
So usually I try to focus on
how problem to solve. Back then
26
00:02:06.549 --> 00:02:10.189
it was just as in the area
of marketing, marketing automation. We had
27
00:02:10.669 --> 00:02:16.020
the market move towards more online,
in digital. So focusing on the problem
28
00:02:16.180 --> 00:02:20.139
is great, as you what they're
trying to deliver, what type of solution
29
00:02:20.219 --> 00:02:23.099
and how the solution will actually drive
value to your end customers and can you
30
00:02:23.219 --> 00:02:27.699
make it ten next better than the
other players? So I think one of
31
00:02:27.780 --> 00:02:30.770
the things just to focus on the
problem and then just don't shy away from
32
00:02:30.849 --> 00:02:35.969
big problems, and I remember that, you know, but there's over time,
33
00:02:36.129 --> 00:02:38.530
over the years, you always find
ways to solve those problem in this
34
00:02:38.689 --> 00:02:44.840
one and just applying these different strategies
and out of that is again like how
35
00:02:44.919 --> 00:02:49.560
can you turn those data points into
into insights and into action? So one
36
00:02:49.639 --> 00:02:53.560
other best practice would just to say, like try to close the loop and
37
00:02:53.639 --> 00:02:57.000
not just do part of the problem. Just try to say that side to
38
00:02:57.039 --> 00:03:00.469
see that you can can really take
it the end to end and just delivering
39
00:03:00.550 --> 00:03:02.349
the most value for a customers.
Yeah, I mean I think that's a
40
00:03:02.469 --> 00:03:07.669
really important point. Often Times companies
come out to solve a problem but it's
41
00:03:07.710 --> 00:03:12.629
not actually the solution for the market
or the customers. It is a solution
42
00:03:12.710 --> 00:03:15.460
to a problem, but it's not
the solution the customers are looking for expecting.
43
00:03:15.580 --> 00:03:19.419
Right, absolutely, I think that
one thing that you don't really control
44
00:03:19.580 --> 00:03:23.979
is obviously the market and timing.
So for my entreprene earlier background, I
45
00:03:23.020 --> 00:03:29.050
would say that you want to consense
when the market is ready and how painful
46
00:03:29.849 --> 00:03:32.490
is the problem, because those things
you can't really control. It's out of
47
00:03:32.569 --> 00:03:37.090
your control your con you can control
what you can focus on and what you
48
00:03:37.210 --> 00:03:42.439
want to deliver. And but those
two elements led me also back then to
49
00:03:42.560 --> 00:03:46.360
kind of go into the marketing space
after being in the financial space, and
50
00:03:46.520 --> 00:03:50.919
also led me to go into this, would say product experience and customer experience
51
00:03:50.919 --> 00:03:54.479
space because, you know, the
the market realization of that is a big
52
00:03:54.639 --> 00:03:59.909
problem to go solve. Just is
happening. So I was fortunate at twice
53
00:04:00.189 --> 00:04:03.069
go through to MNAS and hitting the
the market with the right product at the
54
00:04:03.110 --> 00:04:08.550
right time. And then it was
very, very hard, of course,
55
00:04:08.629 --> 00:04:12.180
to be in competition, but I
think it's competition is a great thing and
56
00:04:12.500 --> 00:04:15.620
we saw that every the the problem
that we were focusing on and the fact
57
00:04:15.699 --> 00:04:20.699
that we build something which is defensible
really helped us, you know, really
58
00:04:20.819 --> 00:04:25.490
scale and deliver the value during the
day. But back then it was real
59
00:04:25.610 --> 00:04:29.930
time personalization to your website and in
reaching data, for example, because you
60
00:04:30.009 --> 00:04:33.769
want to make sure that you're able
to track the right data. The other
61
00:04:33.850 --> 00:04:39.680
challenge would be how can you really
drive insights and then, third how can
62
00:04:39.680 --> 00:04:44.959
you drive results? So these three
elements is something traditionally see companies are struggling
63
00:04:45.000 --> 00:04:47.360
with. They can find how to
focus on the right data or collect the
64
00:04:47.399 --> 00:04:50.519
right data and then driving the right
insights. And then, once you have
65
00:04:50.600 --> 00:04:54.750
the insights, are you able to
take action. Are you able to drive
66
00:04:54.829 --> 00:04:58.069
results? Is kind of the challenge
when you think about the problem. You
67
00:04:58.189 --> 00:05:00.790
want to make sure that you are
able. You're able to drive results,
68
00:05:00.870 --> 00:05:03.069
not just, you know, help
them with one of the elements. In
69
00:05:03.149 --> 00:05:08.819
the second startup, I realized that
you know that collecting data is this something
70
00:05:08.899 --> 00:05:12.779
that is not trivial. Most companies
are not graded at collecting the right data.
71
00:05:13.379 --> 00:05:16.819
So in our solution today, with
gains agepx, we automatically collect,
72
00:05:17.060 --> 00:05:21.490
for example, any user interaction with
any application, Anyui. It was very
73
00:05:21.810 --> 00:05:29.129
aspirational to say we can collect website, web application, mobile and even windows
74
00:05:29.209 --> 00:05:32.089
desktop. We want to collect every
type of user interaction, synthesize that data
75
00:05:32.250 --> 00:05:36.000
and then help that data become part
of the analytics. So the collection of
76
00:05:36.079 --> 00:05:41.319
the data is something that is overlooked
by many solutions. They will they assume
77
00:05:41.600 --> 00:05:44.279
that the company will have a way
to collect the data, but I think
78
00:05:44.319 --> 00:05:46.600
that's a wrong assumption. I think
that you need to help companies to actually
79
00:05:46.720 --> 00:05:50.829
collect the data and to collect the
right data, because bad data in will
80
00:05:50.870 --> 00:05:55.230
not lead to a great outcome as
well. Yeah, that's true. So
81
00:05:56.149 --> 00:06:00.750
in that what's the hardest part of
collecting the data or accessing it as many,
82
00:06:00.829 --> 00:06:03.899
many elements. There's like starting from
performance. Everybody saying like with the
83
00:06:04.180 --> 00:06:08.500
degregate my application performance. So you
need to know how to do it in
84
00:06:08.579 --> 00:06:13.819
a very effective way with minimal to
zero impact on application performance, because if
85
00:06:13.860 --> 00:06:17.769
you're trying to improve user experience but
you degregate performance, you're just gonna get
86
00:06:17.810 --> 00:06:21.730
the wrong results. There's other elements. For example, is volume. You
87
00:06:21.810 --> 00:06:26.250
know, if we're we're tracking today, you know a lot of hundreds of
88
00:06:26.329 --> 00:06:31.449
products and millions of users and billions
of user interaction. So scale is definitely
89
00:06:31.850 --> 00:06:35.439
something to think about. How can
you track that scale and how can you
90
00:06:35.639 --> 00:06:41.040
even build a sustainable business model,
because storing did and posting data is also
91
00:06:41.160 --> 00:06:43.879
costly. So how can you make
it if it very, very effective?
92
00:06:44.160 --> 00:06:48.470
So it's about, you know,
not being intrusive scale and also about security.
93
00:06:48.990 --> 00:06:51.550
You want to make sure that you're
able to make sure that you are
94
00:06:51.629 --> 00:06:56.430
not collecting any PII, not come
out compromise, the GDP are all those
95
00:06:56.470 --> 00:06:59.550
things are very, very important in
these days. So that's also something to
96
00:06:59.589 --> 00:07:01.860
think about. When you're collecting data. Can you make it safe so your
97
00:07:01.939 --> 00:07:05.220
customers not doing any mistakes and just
coming to say hey, we just collected
98
00:07:05.300 --> 00:07:09.860
like DII and it's a bad news
for everybody. So and as well as
99
00:07:09.939 --> 00:07:12.699
security, as we said before,
E. Security is also an element when
100
00:07:12.699 --> 00:07:15.009
you're collecting data other people. We
want to get to that data and the
101
00:07:15.050 --> 00:07:19.009
way you collect data is also another
area that's hackers, for example, will
102
00:07:19.050 --> 00:07:23.769
find very interesting. So you want
to make sure that you aware of the
103
00:07:24.050 --> 00:07:29.009
security implication and make sure that you're
almost as as secured as possible and that
104
00:07:29.449 --> 00:07:32.040
respect. Yeah, definitely. I
mean secured is a big issue today,
105
00:07:32.120 --> 00:07:35.199
of course, and a lot of
challenges with privacy. On the side of
106
00:07:35.279 --> 00:07:39.319
scale, it's really interesting because we've
talked to a lot of companies and you
107
00:07:39.399 --> 00:07:42.680
know, a lot of companies struggle
with the scale and and reality of the
108
00:07:42.759 --> 00:07:45.350
World War and today, where there's
much more data than there was in the
109
00:07:45.430 --> 00:07:48.149
past. But the next layer they
were starting to hear a lot more about
110
00:07:48.350 --> 00:07:51.709
is the real time aspect. So
it's one thing to be able to store
111
00:07:51.709 --> 00:07:55.750
a lot of data and put it
a data warehouse somewhere. It's a whole
112
00:07:55.790 --> 00:07:59.339
other thing to be able to make
decisions based on data the streaming or a
113
00:07:59.579 --> 00:08:03.220
live environment. I'd love to hear
more about kind of how you've dealt with
114
00:08:03.379 --> 00:08:07.220
the the streaming or live data challenges. Yeah, absolutely. I think when
115
00:08:07.259 --> 00:08:11.860
you go into this the two paradigms. One is the big data analytics where
116
00:08:13.170 --> 00:08:18.449
customers expect to see data in seconds, minutes, sometimes hours. The aspiration
117
00:08:18.610 --> 00:08:22.250
is to be in the seconds really
space, I would say. But on
118
00:08:22.370 --> 00:08:26.449
the real time side, because as
a solution sometimes you want to trigger re
119
00:08:26.610 --> 00:08:30.920
time action. It might be just
you know, if you're optimizing your ads,
120
00:08:31.000 --> 00:08:33.000
for example, it might be that
you know you need to send the
121
00:08:33.080 --> 00:08:37.879
bid in the in re time.
Back then, when I dial with financial
122
00:08:37.919 --> 00:08:41.950
market, that's the real time decision
of buy and say, and today the
123
00:08:41.110 --> 00:08:45.830
real time might be just the right
campaign and want to serve for the users.
124
00:08:46.470 --> 00:08:50.110
And really time is in the in
the space, so on the the
125
00:08:50.190 --> 00:08:52.950
range of the millisecond real time.
And what you have to have in with
126
00:08:54.070 --> 00:08:58.620
real time is predictability, because you
have just a minimum window to take action
127
00:08:58.940 --> 00:09:05.059
and it has to be sustainable.
If suddenly one mistake there or one logical
128
00:09:05.100 --> 00:09:09.700
error, suddenly the dire real time
just become seconds. So for monitoring perspective,
129
00:09:09.740 --> 00:09:11.929
everything is green, no, but
nothing actually is not functioning. But
130
00:09:13.450 --> 00:09:18.169
from the business perspective, you're not
delivering the value. So as soon as
131
00:09:18.210 --> 00:09:22.210
you breached that milliseconds ranged into seconds, that is is a breach in your
132
00:09:22.330 --> 00:09:26.639
SLA. So there's many things to
think about when you think about real time
133
00:09:26.200 --> 00:09:31.200
action. How much data can you
analyze in real time? In theory,
134
00:09:31.279 --> 00:09:33.799
for example, if we have unlimited
time, we can take great decisions.
135
00:09:33.840 --> 00:09:37.480
We can scan ton of data in
real time. You can just cannot afford.
136
00:09:37.480 --> 00:09:41.309
You have to optimize for the best
decision within the time so you can
137
00:09:41.389 --> 00:09:45.350
close the little bit and do something
about that, versus, you know,
138
00:09:45.590 --> 00:09:48.990
just you know, spending a lot
of time but the customer is already gone.
139
00:09:48.110 --> 00:09:52.190
So your answer is great, but
there's not going to anyone to listen.
140
00:09:52.789 --> 00:09:54.980
So real time has a lot of
challenges and it starts from how you
141
00:09:56.179 --> 00:10:01.179
model the data, how you process
that, how you guarantee that response time
142
00:10:01.539 --> 00:10:05.820
SLA, for example, the predictability
in that, and how can you increase
143
00:10:05.860 --> 00:10:11.610
all the time the amount of data
you can trigger, you can actually analyze
144
00:10:11.730 --> 00:10:15.330
before you come back? And technology
is always evolving. So obviously using in
145
00:10:15.490 --> 00:10:20.289
memory is is giving us the that's
the fastest speed to analyze. But there's
146
00:10:20.289 --> 00:10:24.240
a lot to do with the way
you model in architecture solution to allow the
147
00:10:24.360 --> 00:10:28.440
best of the kind of allowing for
the best decision to be made within a
148
00:10:28.600 --> 00:10:31.279
very, very short time, and
they're very strict tessellary. So it's about
149
00:10:31.320 --> 00:10:35.590
the modeling the product, you know, the it's about optimizing the size of
150
00:10:35.710 --> 00:10:41.309
your data model, optimizing the execution
rules, you know, if we're,
151
00:10:41.389 --> 00:10:46.070
for example, running twenty, thirtyzero
rules per second in many, many cases,
152
00:10:46.789 --> 00:10:50.980
and and it obviously has to be
smart execution because eventually every user wants
153
00:10:52.019 --> 00:10:54.539
to get the right message of the
right time within the milliseconds, because are
154
00:10:54.539 --> 00:10:58.059
right now, are on the website
or in your product and you have just,
155
00:10:58.259 --> 00:11:01.179
you know, less than five hundred
milliseconds to close them. So that's,
156
00:11:01.299 --> 00:11:05.009
though, the the what we live
in and the fact that you need
157
00:11:05.090 --> 00:11:07.970
to think about the way you run
the you build a model. The run
158
00:11:07.090 --> 00:11:13.570
time is critical and many engineers just
are usually are kind of focusing on area
159
00:11:13.730 --> 00:11:18.049
when seconds is fine, you know, even three to five seconds everything.
160
00:11:18.129 --> 00:11:22.519
So it's to become an engineer that
knows how to deal with real time.
161
00:11:22.559 --> 00:11:24.840
You need to can shift the way
you think about everything you do, starting
162
00:11:24.879 --> 00:11:30.360
from coding, the size of object
you're building, the way you understand multi
163
00:11:30.480 --> 00:11:35.750
threading? All those elements we actually
learning university but scarcely apply in our day
164
00:11:35.789 --> 00:11:39.230
to day because most application are fine
with, you know, two or three
165
00:11:39.269 --> 00:11:43.870
seconds and your code can be very
inefficient and will still work, as opposed
166
00:11:43.909 --> 00:11:46.750
to real time when, if your
code is not in is inefficient, you're
167
00:11:46.750 --> 00:11:48.700
not going to be able to deliver
that at scale. Yeah, that makes
168
00:11:48.740 --> 00:11:52.779
sense. It sounds like what you're
describing a lot of times in the real
169
00:11:52.940 --> 00:11:58.139
time environment is really user experience.
So it's really the customization of how you're
170
00:11:58.179 --> 00:12:01.809
delivering to the user. And it's
one thing on a website with some content
171
00:12:01.970 --> 00:12:05.330
to serve up some better content right
past content, but it's a whole other
172
00:12:05.370 --> 00:12:09.289
thing when you're talking about a full
product experience. Right. Can you talk
173
00:12:09.330 --> 00:12:11.529
a little bit about about that and
where you see the evolution too, of
174
00:12:11.809 --> 00:12:16.080
where we're at, you know,
as far as what expectations are of the
175
00:12:16.159 --> 00:12:18.720
customer. So the expectation is again
to be able to do the the three
176
00:12:18.799 --> 00:12:22.440
things like one, to to understand
what the customer is because then it will
177
00:12:22.480 --> 00:12:26.759
drive the best decisions. If you're
limited in the what elements, what type
178
00:12:26.759 --> 00:12:31.470
of data input you have, then
the output won't be as effective. So
179
00:12:31.750 --> 00:12:35.309
number one is to really make sure
that we are collecting in real time as
180
00:12:35.389 --> 00:12:37.350
much data as possible. So we
want to understand what the customer is.
181
00:12:37.750 --> 00:12:41.350
Why did they do in the past
three months, for example, should so
182
00:12:41.549 --> 00:12:45.340
there's very simple use cases. I'm
building new features. How can I drive
183
00:12:45.379 --> 00:12:48.500
you to know those features and how
can I make sure that it's relevant for
184
00:12:48.620 --> 00:12:52.820
you? So I need to know
did you use it before, what type
185
00:12:52.820 --> 00:12:56.580
of usage pattern you have? All
that type is based on the decision,
186
00:12:56.740 --> 00:12:58.850
based on the collection of data that
I can do. It means that I
187
00:12:58.970 --> 00:13:03.610
want to understand, for example,
feature usage. I want to understand who
188
00:13:03.610 --> 00:13:07.090
you are as a persona this has
to be part of the model that you're
189
00:13:07.129 --> 00:13:09.970
building. So, for example,
in our solution we translate and call it
190
00:13:11.129 --> 00:13:13.440
every click they use a does to
a feature emodule. Also, we have
191
00:13:13.519 --> 00:13:16.720
better context of what you're doing right
now, because I don't want to recommend
192
00:13:16.759 --> 00:13:20.120
a feature you just clicked on.
You just went there and I'm like,
193
00:13:20.240 --> 00:13:22.919
Hey, learn about this, like
I already did that a second ago.
194
00:13:24.120 --> 00:13:28.830
So ast type of message, because
we know that today, as many ways
195
00:13:28.990 --> 00:13:33.710
to consume content, as many ways
to users choose different ways to use your
196
00:13:33.789 --> 00:13:37.269
product. So what would be the
best messaging? So the triggering of that
197
00:13:37.429 --> 00:13:43.059
message, it has to come to
come in into the to the product in
198
00:13:43.139 --> 00:13:46.259
real time, might be very different. So what you have to do is
199
00:13:46.340 --> 00:13:48.940
a top of the triggering rules,
you want to make sure that you have
200
00:13:50.419 --> 00:13:54.179
prioritization rules and you have some fadling
rules to make sure that you're not interfering
201
00:13:54.259 --> 00:13:58.769
with the user interaction, just helping
the user but not blocking the users.
202
00:13:58.850 --> 00:14:01.809
And if there's collision of different type
of messages, they might be all be
203
00:14:01.970 --> 00:14:05.409
relevant, but the expectation is the
system to be smart enough to choose the
204
00:14:05.490 --> 00:14:09.960
right one. So in some cases
we let you know the users prioritize,
205
00:14:11.039 --> 00:14:16.159
but the vision is to have the
system decide what's best, because you might
206
00:14:16.200 --> 00:14:20.200
qualify to more than one elements.
And the last thing is, I the
207
00:14:20.240 --> 00:14:22.720
expectation is to now that we are, you know, collecting so, so
208
00:14:22.799 --> 00:14:26.789
much data and a click of a
button today we can map your entire product
209
00:14:26.909 --> 00:14:31.710
and suddenly you can have visibility to
five hundred features. Were in the past
210
00:14:31.830 --> 00:14:35.429
you saw, for example, ten
features because you have your engineer collect and
211
00:14:35.549 --> 00:14:39.220
instrument ten features. But now,
suddenly, with automation, we can surface
212
00:14:39.220 --> 00:14:41.299
all your features, as a lot
of features to instrument. How can they
213
00:14:41.620 --> 00:14:45.779
make you smart about the data?
which features you should look at? What
214
00:14:45.899 --> 00:14:50.500
type of KPIS and patterns? Cause
as users we want to get the insights,
215
00:14:50.779 --> 00:14:52.610
not necessarily all the reporting. Right
you go. We don't want to
216
00:14:52.610 --> 00:14:56.850
have yet more data. We want
to be smart about the data. So
217
00:14:56.009 --> 00:15:01.009
Are our vision. Our goal is
to help you distill the right set of
218
00:15:01.090 --> 00:15:03.250
data so you can really be better
at your job. And that's it.
219
00:15:03.409 --> 00:15:07.840
That's a challenge because in data you
know that you need to have the right
220
00:15:07.919 --> 00:15:11.759
context and you need to know what
is the right thing to surface to a
221
00:15:11.919 --> 00:15:16.279
person to assist them with decisioning.
So, for example, what is the
222
00:15:16.399 --> 00:15:22.509
pattern of a good user behavior?
What is the pattern of a customer at
223
00:15:22.590 --> 00:15:26.309
risk? This type of things we
aspire to build using machine learning. Just
224
00:15:26.470 --> 00:15:30.190
understand how can I surface things to
you, because as human being, you
225
00:15:30.269 --> 00:15:33.870
cannot just digest so much data.
You have to get this assistance so I
226
00:15:33.950 --> 00:15:39.620
can point you towards a normally they
can point you towards like customers at most
227
00:15:39.700 --> 00:15:43.820
risk that will impact the most revenue
for you. Those type of decisions today
228
00:15:43.820 --> 00:15:48.580
are manual. Most companies are not
good collecting data. They usually collect what
229
00:15:48.700 --> 00:15:50.850
they can and then they search for
based on the data that they have,
230
00:15:52.090 --> 00:15:54.850
but they might be missing, you
know, around seventy percent of the potentially
231
00:15:56.009 --> 00:15:58.210
that they should look at. So
there's a lot of challenges and we are
232
00:15:58.250 --> 00:16:00.649
trying to kind of address all of
them, from collection to insights to drive
233
00:16:00.690 --> 00:16:06.159
action, and our vision is to
start to harness machine learning to help us
234
00:16:07.240 --> 00:16:11.480
be kind of drive better decisions for
our end users. Yeah, that's interesting.
235
00:16:11.559 --> 00:16:15.919
So how much are you relying on
machine learning models versus let's say,
236
00:16:15.000 --> 00:16:18.830
business rules or just simple regressions?
You Know How partant is that in your
237
00:16:18.870 --> 00:16:23.309
business? So today I would say
that it's still ninety percent like business rules,
238
00:16:23.389 --> 00:16:26.269
but more complex one as well,
and some of them are automatical,
239
00:16:26.509 --> 00:16:30.750
automatic and based on score, even
scoring. Today I think it's still eventually
240
00:16:30.750 --> 00:16:34.419
business rules, even if you build
very complex scoring, it's start to tat
241
00:16:34.539 --> 00:16:38.019
that. You were, you know, some of the kind of statistical modeling,
242
00:16:38.100 --> 00:16:41.659
but still it's pretty basic, cause
you are not the weights you're putting
243
00:16:41.700 --> 00:16:47.690
on, not necessarily the more gut
feeling or your experience versus they will data
244
00:16:47.730 --> 00:16:52.850
analysis that you should be doing.
Because so we basically want to do to
245
00:16:52.009 --> 00:16:56.889
give our customers the way to run
multiple models based on the data. What
246
00:16:57.009 --> 00:17:02.320
we sold for them is actually the
collection and the fact that we get clean
247
00:17:02.559 --> 00:17:06.559
data in if we are saying this
is a feature use, this is actually
248
00:17:06.559 --> 00:17:10.079
a feature. You don't need to
worry about the what we call the overfitting
249
00:17:10.160 --> 00:17:12.200
on too much, you know,
junk data. So the cleanness of the
250
00:17:12.240 --> 00:17:15.470
data is the one thing that really
helps us build better models. We do
251
00:17:15.670 --> 00:17:19.109
think that every company should have way
more than one model. When you building
252
00:17:19.150 --> 00:17:22.230
predictive models, the way I see
it, you need to slice the customer
253
00:17:22.309 --> 00:17:27.150
journey to, you know, the
different stages and build model to optimize for
254
00:17:27.269 --> 00:17:32.099
each. So if you try to
have any machine learning to solve a specific
255
00:17:32.220 --> 00:17:36.220
problem, I would say the you
probably going to have multiple models. And
256
00:17:36.339 --> 00:17:38.339
what we want to do is give
you the option to yes, tune the
257
00:17:38.420 --> 00:17:42.779
molds, run different molds and see
what actually performs best based on the goals
258
00:17:44.019 --> 00:17:47.490
that you have. For example,
if it's trial conversion and you're trying to
259
00:17:47.529 --> 00:17:48.970
build them all to say what would
be the best, you know, feature
260
00:17:49.009 --> 00:17:53.609
to promote or type of campaigns to
run, there's many model you can run.
261
00:17:53.809 --> 00:17:56.529
We are going to allow you to
run multiple and then you're going to
262
00:17:56.569 --> 00:18:00.720
see the result and you're going to
see their accuracy and decide what's best to
263
00:18:00.799 --> 00:18:04.400
you. And for onboarding, adoption, retention, all those areas are going
264
00:18:04.400 --> 00:18:08.920
to have different models because you might
be even different persona different segments of companies.
265
00:18:10.039 --> 00:18:11.240
One of the biggest mistake I've done, I've seen is, like you
266
00:18:11.319 --> 00:18:17.150
know, companies are building a holistic
adoption score one for all, and I
267
00:18:17.309 --> 00:18:19.789
think you need you need at least
twenty, because you have five person as,
268
00:18:19.789 --> 00:18:25.109
you have four four different cycles,
stages in the customer life cycle and
269
00:18:25.190 --> 00:18:30.619
you have did different inputs fir stage. So I think the what we're trying
270
00:18:30.660 --> 00:18:33.180
to do is to make that very
accessible for you to build models, run
271
00:18:33.259 --> 00:18:36.900
them, even if you're not the
data scientists, and, based on the
272
00:18:36.940 --> 00:18:40.740
accuracy, just decide what is the
best input and what is the best algorithm
273
00:18:40.779 --> 00:18:42.609
to use. That makes sense.
You know, with this, when you're
274
00:18:42.650 --> 00:18:45.250
when you're working with machine learning models, what it would you say is the
275
00:18:45.369 --> 00:18:48.730
number one challenge? You know,
sort of sort of working with AI machine
276
00:18:48.769 --> 00:18:52.450
learning. You know it's exploring everywhere, obviously, but there's still a lot
277
00:18:52.490 --> 00:18:56.119
of challenges in the road in the
roadway. You know what, would you
278
00:18:56.119 --> 00:19:00.640
say? Some of the significant ones
are absolutely I think if I look back
279
00:19:00.680 --> 00:19:04.480
almost would say a to ten years
ago when I first time, I stopped
280
00:19:04.519 --> 00:19:08.799
using machine learning. Back then I've
had some libraries, but nobody, you
281
00:19:08.880 --> 00:19:12.549
know, help you with collecting data. Nobody help you with running, for
282
00:19:12.630 --> 00:19:15.990
example, all the learning and testing. Basically you need to have this,
283
00:19:17.190 --> 00:19:21.309
the the whole process, to run
that. So you needed to build those
284
00:19:21.390 --> 00:19:26.019
data pipelines to go and basically having
the training set and testing set and so
285
00:19:26.539 --> 00:19:30.779
beyond even understand the whole concept.
It was just, you know, patching
286
00:19:30.819 --> 00:19:34.700
those solutions. And let's say you've
passed all those obstacles. Now you have
287
00:19:34.900 --> 00:19:41.329
a model that you want to stop
running data against it. What will help
288
00:19:41.369 --> 00:19:44.369
you serve that like just to make
it like? Cause, again, we
289
00:19:44.450 --> 00:19:48.930
were in the we built a recommendation
count recommendation engine. So for us is
290
00:19:48.130 --> 00:19:52.880
we need to ask the model and
get the response and render content based on
291
00:19:52.039 --> 00:19:56.680
that. So all these were you
had to build versus use any type of
292
00:19:56.759 --> 00:20:00.599
solution. And over time, we're
saying more and more problems are being sold
293
00:20:00.720 --> 00:20:03.440
and I think now in the last
couple of years, you're seeing how all
294
00:20:03.480 --> 00:20:07.829
the big vendors are now really helping
you run models at scale with a lot
295
00:20:07.829 --> 00:20:14.150
of data and even, in some
points, helping you serving and basically hitting
296
00:20:14.190 --> 00:20:18.829
these models and getting responses. So
a lot of obstacles are going down now.
297
00:20:18.990 --> 00:20:21.430
It's the time that we are think, you know, the the challenge
298
00:20:21.430 --> 00:20:22.859
would be like, how do you
choose the right data? How do you
299
00:20:22.940 --> 00:20:26.380
avoid overfitting? How do you avoid, you know, just being biased because
300
00:20:26.460 --> 00:20:30.500
you're bringing a model data set.
You're able to be to bring it not
301
00:20:30.579 --> 00:20:33.859
might not have the answer to begin
with. So common sense is always good
302
00:20:34.220 --> 00:20:37.170
and I've seen no, you know, Super Smart People. They just like
303
00:20:37.369 --> 00:20:41.609
the the the data source they use. They did not question the data source,
304
00:20:41.650 --> 00:20:44.769
they did not question their quality,
you know, they kind of looked
305
00:20:44.769 --> 00:20:45.769
at it as as kind of an
end game. This is the day that
306
00:20:45.769 --> 00:20:49.799
I have as opposed to know,
you should be exploring a lot of different
307
00:20:49.799 --> 00:20:53.079
data sets. Maybe the answer is
not there. That's a very common mistake.
308
00:20:53.160 --> 00:20:57.559
And second challenges again to understand and
make sense of it. And again
309
00:20:57.599 --> 00:21:02.720
sometimes you're getting the score and nobody
at instand why it is. So eitherly
310
00:21:02.720 --> 00:21:04.869
choose to trust it or not trust
it, and I think that if you
311
00:21:06.029 --> 00:21:10.549
break down two different scores, different
models, you'll be able to assess better
312
00:21:10.670 --> 00:21:12.990
that. This is why you're getting
this score as opposed to very high level
313
00:21:14.069 --> 00:21:18.430
kind of predictive score. And last
I would say it should be designed to
314
00:21:18.660 --> 00:21:22.779
assist your decision, not replace your
decision. Is always going to be a
315
00:21:22.819 --> 00:21:27.140
human factor that will be involved just
to understand and tune and feed the algorithm
316
00:21:27.140 --> 00:21:32.619
and decide to rerun the model.
By the way, there's another element that
317
00:21:32.779 --> 00:21:37.369
is very common in financial market.
There rebuilding the model almost nightly, and
318
00:21:37.529 --> 00:21:40.690
in the in the bit to be
industry, they build one model per year.
319
00:21:40.809 --> 00:21:44.170
It just you know, you basically
are. You have to understand that
320
00:21:44.289 --> 00:21:47.799
building models and very fine is somebody
that you need to do more frequently.
321
00:21:48.480 --> 00:21:51.400
Maybe not every twenty four hours,
but you know, at least waterly,
322
00:21:51.440 --> 00:21:53.640
at least monthly, and just morning
to the quality of the service you're getting
323
00:21:53.759 --> 00:21:56.519
from that. But I feel like
we are very getting, very, very
324
00:21:56.640 --> 00:22:00.400
close to that and the market is
ready to start understanding and using it,
325
00:22:00.430 --> 00:22:03.829
as opposed to the past that people
thought it's going to replace human beings.
326
00:22:03.869 --> 00:22:07.910
And I don't trust it and I
see a number and I can't make sense
327
00:22:07.990 --> 00:22:10.430
of it. So I think it's
going to be our future and you see
328
00:22:10.509 --> 00:22:12.950
more and more people are getting used
to assistance and we use you know,
329
00:22:14.349 --> 00:22:17.859
Google, Alexa and all those suff
that today's you know, the next generation
330
00:22:17.940 --> 00:22:23.259
is obviously completely trusting you know ai
technologies to to driving decisions without any question.
331
00:22:23.299 --> 00:22:26.940
Yeah, and I think Covid,
you know, changed a lot of
332
00:22:26.059 --> 00:22:29.500
things, obviously, and a lot
of ways. But I can tell you
333
00:22:29.660 --> 00:22:33.369
that with us and what we look
at in our customers, we saw a
334
00:22:33.529 --> 00:22:37.690
huge change in they're thinking. Or
because they experience Covid, they realize what
335
00:22:37.769 --> 00:22:41.009
you were just saying, which is, Oh, that model that we rebuild
336
00:22:41.049 --> 00:22:42.890
once a year, maybe we need
to do it more frequently, because they
337
00:22:42.890 --> 00:22:47.519
were totally at a loss, you
know, when covid hit, for what
338
00:22:47.680 --> 00:22:51.079
to do next or how to really
customize the experience for their customers. Right.
339
00:22:51.119 --> 00:22:55.480
Yeah, in many cases what happened
is like in most models, you
340
00:22:55.640 --> 00:23:00.190
basically looking backwards on data and driving
forward, and when covid hits, you're
341
00:23:00.190 --> 00:23:04.109
not going to see the tree understanding
that you need to refresh your model and
342
00:23:04.390 --> 00:23:08.789
move towards, you know, shorter
cycles and there's a whole element of new
343
00:23:08.950 --> 00:23:14.420
networks. But I think, but
I think in the classic data science you
344
00:23:14.740 --> 00:23:18.980
want to make shorter windows of time
to analyze the data so you're getting you're
345
00:23:18.059 --> 00:23:22.180
able to at least react. The
other thing is is now it's going to
346
00:23:22.220 --> 00:23:27.289
become much cheaper to use data science. I think it's more accessible. I
347
00:23:27.410 --> 00:23:33.089
think still I see companies that have
only say companies have the luxury and budget
348
00:23:33.170 --> 00:23:37.690
to have data science team and to
use, you know, those type of
349
00:23:37.970 --> 00:23:41.559
processes. But I think in the
future it's going to be available for for
350
00:23:41.720 --> 00:23:45.119
any size company to start utilizing machine
learning because because there's a lot of things
351
00:23:45.200 --> 00:23:48.279
to do. And the second thing
I see a lot is, again the
352
00:23:48.680 --> 00:23:52.079
scoring model. At Marquetto we had
at leade scoring, and most of the
353
00:23:52.079 --> 00:23:56.750
lead scoring I've seen by customers is
just that feeling of marketers. They say,
354
00:23:56.789 --> 00:23:59.630
if you clicked on this website,
you download this white paper, you
355
00:23:59.710 --> 00:24:03.549
qualifying like that's and even they might
enrich the data, but to really understand
356
00:24:03.670 --> 00:24:07.710
what behavior drives the best conversion,
you have to apply a machine and you
357
00:24:07.789 --> 00:24:12.460
have to apply much stronger dataalysis,
because you have millions of website visitors,
358
00:24:12.500 --> 00:24:17.099
for example, or millions of users, and there's no way you will be
359
00:24:17.180 --> 00:24:21.140
able to with your gut feeling to
score that. Today we're seeing with customer
360
00:24:21.220 --> 00:24:25.930
health score, we're with bullock adoption
scored as many scores that the data science
361
00:24:25.970 --> 00:24:30.609
teams are trying to do. So
when it's a data science team they tend
362
00:24:30.650 --> 00:24:33.450
to know exactly what they're doing in
the score tends to be pretty accurate.
363
00:24:33.849 --> 00:24:37.130
But not every company can afford a
full data science team and all that kind
364
00:24:37.130 --> 00:24:41.960
of data aggregation and processing, and
I think that's going to be something that's
365
00:24:41.960 --> 00:24:45.440
going to be more affordable to other
companies. Yeah, it's to playing a
366
00:24:45.519 --> 00:24:48.400
huge Doman for more data science talent, you know, as we move forward
367
00:24:48.440 --> 00:24:53.109
in the world hoping to make more
data driven decisions. Right. What advice
368
00:24:53.190 --> 00:24:56.990
would you give to someone seeking your
career in data science? You know,
369
00:24:56.069 --> 00:25:00.710
based on your experience, I think
that in data science and again, I
370
00:25:00.869 --> 00:25:04.910
always looked at data as kind of
what is the oil to drive business results.
371
00:25:06.109 --> 00:25:10.779
In marketing again, was the input
for personalization, for better statistics,
372
00:25:10.859 --> 00:25:14.579
for better insights about your customer intent. I think it's important for everyone,
373
00:25:14.660 --> 00:25:18.259
data science or engineering, just to
really understand and also care about the product
374
00:25:18.299 --> 00:25:22.170
and in the outcome. Just as
a general advice, but I think you
375
00:25:22.289 --> 00:25:26.089
know, it's becoming a very needed
and required talent in every company. So
376
00:25:26.369 --> 00:25:32.170
learning all the tools, applying this
different models and learning all the time is
377
00:25:32.210 --> 00:25:34.130
going to be very, very useful
because I think again, the market is
378
00:25:34.250 --> 00:25:38.880
now becoming ready for a machine learning
driven type of solutions, and I've seen
379
00:25:40.000 --> 00:25:42.960
those solutions actually around two thousand and
fourteen, like six seven years ago,
380
00:25:44.559 --> 00:25:48.680
there were a lot of machine learning
driven solutions, like lead predictively scoring in
381
00:25:48.799 --> 00:25:52.430
others that were way ahead of that
time and I think also they lack some
382
00:25:52.549 --> 00:25:56.950
of the technologies that exist today where
now they're very, very successful. But
383
00:25:56.069 --> 00:25:59.309
I think in two thousand and twenty
one, two thousand and twenty two,
384
00:25:59.430 --> 00:26:03.980
machine learning is becoming very except acceptable
and accessible and the need for data sciences
385
00:26:04.099 --> 00:26:11.140
to really be able to run those
type of motions are increasingly higher as well.
386
00:26:11.500 --> 00:26:15.339
I mean it's interesting that now,
as this whole industry revolves, what
387
00:26:15.460 --> 00:26:19.049
we're realizing is that this concept of
the growth mindset or learning or continuous learning,
388
00:26:21.250 --> 00:26:25.529
when you take people who are most
comfortable in math and technology and then
389
00:26:25.569 --> 00:26:29.049
you put them into an environment which
you have to or right to be in
390
00:26:29.089 --> 00:26:32.210
that world, but you put them
an environment where they're constantly have to having
391
00:26:32.250 --> 00:26:37.079
to evolve and change and and learn
about the industry and what's happening because,
392
00:26:37.079 --> 00:26:40.319
as you said, a few years
makes a huge difference, and what we
393
00:26:40.440 --> 00:26:42.599
can do and what the costs are
doing. That right, absolutely, and
394
00:26:42.640 --> 00:26:45.720
I think that they need two questions. A lot of the kind of facts
395
00:26:45.920 --> 00:26:49.549
to begin with, because again they
might just get a problem to solve and
396
00:26:49.589 --> 00:26:52.430
the data set, but they need
to question, is this really the data
397
00:26:52.509 --> 00:26:56.670
set? This is all we have? Can we actually exploded? We explore
398
00:26:56.670 --> 00:27:00.309
other things before you get to this
more optimization, because again, if with
399
00:27:00.549 --> 00:27:03.660
bed oil you can't really have a
great, you know, fast moving car.
400
00:27:03.900 --> 00:27:07.299
So it's that's that's something that you
want to make sure that you're always
401
00:27:07.299 --> 00:27:11.579
learning. And they need to also
understand the space, the problem that they're
402
00:27:11.660 --> 00:27:15.049
solving, because otherwise they would not
have these kind of curiosity questions. Again,
403
00:27:15.130 --> 00:27:19.130
I was in the marketing space or
in the financial space, understanding the
404
00:27:19.250 --> 00:27:23.809
paradigms there really helps you as data
sciences to understand this might be relevant data,
405
00:27:23.849 --> 00:27:27.690
this might be, you know,
less relevant data, and be more
406
00:27:27.730 --> 00:27:32.920
creative there, because eventually it requires
you to think what would be the best
407
00:27:32.920 --> 00:27:36.079
input to your models and why would
be the best models to fit. And
408
00:27:36.480 --> 00:27:40.880
again, without knowing the entire spectrum
of what you're solving, you might just
409
00:27:41.000 --> 00:27:44.390
looking under the fleshlight. Is supposed
to look where you know the data is
410
00:27:44.430 --> 00:27:48.750
really exists. Words representative. Yeah, that makes sense. Well, thank
411
00:27:48.789 --> 00:27:51.470
you so much for being on the
show today. You know, we really
412
00:27:51.509 --> 00:27:53.789
appreciate your time. Where can the
community find you? What's the best way
413
00:27:53.789 --> 00:27:57.269
to get in touch with you?
So I'm on Linkedin. My against it's
414
00:27:57.309 --> 00:28:03.019
Mickey alone on Linkedin at gain site
and my email is actually micky at gainsayecom.
415
00:28:03.140 --> 00:28:07.460
So feel free to reach out and
happy to advice. Also, we
416
00:28:07.660 --> 00:28:12.089
have a lot of content with regards
to analytic sport and it takes customer aalytics.
417
00:28:12.170 --> 00:28:15.650
So happy to if you have any
question around that, happy to take
418
00:28:15.690 --> 00:28:18.529
that as well. Okay, Great. Well, thank you. And at
419
00:28:18.569 --> 00:28:22.650
the data standard we are building a
community of data enthusiasts and data thought leaders.
420
00:28:22.690 --> 00:28:26.609
We look forward to engaging with you
on our website, data standard Ioh,
421
00:28:26.009 --> 00:28:30.599
where we regularly post a wide range
of podcasts with data executives and offer
422
00:28:30.680 --> 00:28:36.319
regular networking opportunities. For more information, visit the website. Thank you so
423
00:28:36.440 --> 00:28:37.680
much for joining us today. Thank
you for having me
1
00:00:10.349 --> 00:00:13.789
Hi, welcome to the data standard
and I'm your guest host for today.
2
00:00:13.869 --> 00:00:16.789
Give you in Ubon. On our
show today we have Mickey along, the
3
00:00:16.910 --> 00:00:20.350
CTEO and founder, at gain site. Today we're going to be speaking about
4
00:00:20.750 --> 00:00:24.460
using data for marketing animation. Welcome
to the show. Hi, thank you
5
00:00:24.539 --> 00:00:28.660
for having me. Where are you
located right now? I'm from South Bay.
6
00:00:28.940 --> 00:00:33.659
Moved from Tel Aviv, Israel,
eight years ago. It's part of
7
00:00:34.539 --> 00:00:40.289
building a local presence to a former
start up. Actually, I founded nice
8
00:00:40.450 --> 00:00:43.049
and tell us a little bit about
your background. I'd love to hear more.
9
00:00:43.170 --> 00:00:46.689
Awesome. So my Baker is kind
of from computer science and also did
10
00:00:46.770 --> 00:00:51.799
my MBA degree, but started coding
at age thirteen or less. Kind of
11
00:00:51.920 --> 00:00:57.520
like Geeky with regards to computer science
and building actually software and products. I've
12
00:00:57.640 --> 00:01:03.629
dealt couple of years with retaime ONA
lyrics in the financial markets and then also
13
00:01:04.469 --> 00:01:10.989
started the two companies. One was
about website, web and brand experience.
14
00:01:11.629 --> 00:01:15.989
That kind of applied the same elements
of taking decision in real time and delivering
15
00:01:17.030 --> 00:01:23.700
a better experience for customers, and
now recently co found gains at Px,
16
00:01:23.819 --> 00:01:26.819
the product experience platform. That is
the same. Is just dealing with a
17
00:01:26.900 --> 00:01:30.620
lot of data, taking the those
insights, driving results and just delivering better
18
00:01:30.659 --> 00:01:34.489
customer experience in the end. Yeah, it so it sounds like you've done
19
00:01:34.530 --> 00:01:38.689
some similar things over the years.
Do you have best practices that you would
20
00:01:38.769 --> 00:01:45.290
share that you think you've learned over
out of these different iterations of problems?
21
00:01:45.609 --> 00:01:49.480
It's always about about the data in
the end, it's and you want to
22
00:01:49.560 --> 00:01:53.640
make sure that you're driving outcome for
your customers. So, you know,
23
00:01:53.719 --> 00:01:57.040
one of the best practices is obviously
always focus all the on the problem and,
24
00:01:57.400 --> 00:02:01.150
as they always say, try to
build something that is en x better.
25
00:02:02.109 --> 00:02:06.510
So usually I try to focus on
how problem to solve. Back then
26
00:02:06.549 --> 00:02:10.189
it was just as in the area
of marketing, marketing automation. We had
27
00:02:10.669 --> 00:02:16.020
the market move towards more online,
in digital. So focusing on the problem
28
00:02:16.180 --> 00:02:20.139
is great, as you what they're
trying to deliver, what type of solution
29
00:02:20.219 --> 00:02:23.099
and how the solution will actually drive
value to your end customers and can you
30
00:02:23.219 --> 00:02:27.699
make it ten next better than the
other players? So I think one of
31
00:02:27.780 --> 00:02:30.770
the things just to focus on the
problem and then just don't shy away from
32
00:02:30.849 --> 00:02:35.969
big problems, and I remember that, you know, but there's over time,
33
00:02:36.129 --> 00:02:38.530
over the years, you always find
ways to solve those problem in this
34
00:02:38.689 --> 00:02:44.840
one and just applying these different strategies
and out of that is again like how
35
00:02:44.919 --> 00:02:49.560
can you turn those data points into
into insights and into action? So one
36
00:02:49.639 --> 00:02:53.560
other best practice would just to say, like try to close the loop and
37
00:02:53.639 --> 00:02:57.000
not just do part of the problem. Just try to say that side to
38
00:02:57.039 --> 00:03:00.469
see that you can can really take
it the end to end and just delivering
39
00:03:00.550 --> 00:03:02.349
the most value for a customers.
Yeah, I mean I think that's a
40
00:03:02.469 --> 00:03:07.669
really important point. Often Times companies
come out to solve a problem but it's
41
00:03:07.710 --> 00:03:12.629
not actually the solution for the market
or the customers. It is a solution
42
00:03:12.710 --> 00:03:15.460
to a problem, but it's not
the solution the customers are looking for expecting.
43
00:03:15.580 --> 00:03:19.419
Right, absolutely, I think that
one thing that you don't really control
44
00:03:19.580 --> 00:03:23.979
is obviously the market and timing.
So for my entreprene earlier background, I
45
00:03:23.020 --> 00:03:29.050
would say that you want to consense
when the market is ready and how painful
46
00:03:29.849 --> 00:03:32.490
is the problem, because those things
you can't really control. It's out of
47
00:03:32.569 --> 00:03:37.090
your control your con you can control
what you can focus on and what you
48
00:03:37.210 --> 00:03:42.439
want to deliver. And but those
two elements led me also back then to
49
00:03:42.560 --> 00:03:46.360
kind of go into the marketing space
after being in the financial space, and
50
00:03:46.520 --> 00:03:50.919
also led me to go into this, would say product experience and customer experience
51
00:03:50.919 --> 00:03:54.479
space because, you know, the
the market realization of that is a big
52
00:03:54.639 --> 00:03:59.909
problem to go solve. Just is
happening. So I was fortunate at twice
53
00:04:00.189 --> 00:04:03.069
go through to MNAS and hitting the
the market with the right product at the
54
00:04:03.110 --> 00:04:08.550
right time. And then it was
very, very hard, of course,
55
00:04:08.629 --> 00:04:12.180
to be in competition, but I
think it's competition is a great thing and
56
00:04:12.500 --> 00:04:15.620
we saw that every the the problem
that we were focusing on and the fact
57
00:04:15.699 --> 00:04:20.699
that we build something which is defensible
really helped us, you know, really
58
00:04:20.819 --> 00:04:25.490
scale and deliver the value during the
day. But back then it was real
59
00:04:25.610 --> 00:04:29.930
time personalization to your website and in
reaching data, for example, because you
60
00:04:30.009 --> 00:04:33.769
want to make sure that you're able
to track the right data. The other
61
00:04:33.850 --> 00:04:39.680
challenge would be how can you really
drive insights and then, third how can
62
00:04:39.680 --> 00:04:44.959
you drive results? So these three
elements is something traditionally see companies are struggling
63
00:04:45.000 --> 00:04:47.360
with. They can find how to
focus on the right data or collect the
64
00:04:47.399 --> 00:04:50.519
right data and then driving the right
insights. And then, once you have
65
00:04:50.600 --> 00:04:54.750
the insights, are you able to
take action. Are you able to drive
66
00:04:54.829 --> 00:04:58.069
results? Is kind of the challenge
when you think about the problem. You
67
00:04:58.189 --> 00:05:00.790
want to make sure that you are
able. You're able to drive results,
68
00:05:00.870 --> 00:05:03.069
not just, you know, help
them with one of the elements. In
69
00:05:03.149 --> 00:05:08.819
the second startup, I realized that
you know that collecting data is this something
70
00:05:08.899 --> 00:05:12.779
that is not trivial. Most companies
are not graded at collecting the right data.
71
00:05:13.379 --> 00:05:16.819
So in our solution today, with
gains agepx, we automatically collect,
72
00:05:17.060 --> 00:05:21.490
for example, any user interaction with
any application, Anyui. It was very
73
00:05:21.810 --> 00:05:29.129
aspirational to say we can collect website, web application, mobile and even windows
74
00:05:29.209 --> 00:05:32.089
desktop. We want to collect every
type of user interaction, synthesize that data
75
00:05:32.250 --> 00:05:36.000
and then help that data become part
of the analytics. So the collection of
76
00:05:36.079 --> 00:05:41.319
the data is something that is overlooked
by many solutions. They will they assume
77
00:05:41.600 --> 00:05:44.279
that the company will have a way
to collect the data, but I think
78
00:05:44.319 --> 00:05:46.600
that's a wrong assumption. I think
that you need to help companies to actually
79
00:05:46.720 --> 00:05:50.829
collect the data and to collect the
right data, because bad data in will
80
00:05:50.870 --> 00:05:55.230
not lead to a great outcome as
well. Yeah, that's true. So
81
00:05:56.149 --> 00:06:00.750
in that what's the hardest part of
collecting the data or accessing it as many,
82
00:06:00.829 --> 00:06:03.899
many elements. There's like starting from
performance. Everybody saying like with the
83
00:06:04.180 --> 00:06:08.500
degregate my application performance. So you
need to know how to do it in
84
00:06:08.579 --> 00:06:13.819
a very effective way with minimal to
zero impact on application performance, because if
85
00:06:13.860 --> 00:06:17.769
you're trying to improve user experience but
you degregate performance, you're just gonna get
86
00:06:17.810 --> 00:06:21.730
the wrong results. There's other elements. For example, is volume. You
87
00:06:21.810 --> 00:06:26.250
know, if we're we're tracking today, you know a lot of hundreds of
88
00:06:26.329 --> 00:06:31.449
products and millions of users and billions
of user interaction. So scale is definitely
89
00:06:31.850 --> 00:06:35.439
something to think about. How can
you track that scale and how can you
90
00:06:35.639 --> 00:06:41.040
even build a sustainable business model,
because storing did and posting data is also
91
00:06:41.160 --> 00:06:43.879
costly. So how can you make
it if it very, very effective?
92
00:06:44.160 --> 00:06:48.470
So it's about, you know,
not being intrusive scale and also about security.
93
00:06:48.990 --> 00:06:51.550
You want to make sure that you're
able to make sure that you are
94
00:06:51.629 --> 00:06:56.430
not collecting any PII, not come
out compromise, the GDP are all those
95
00:06:56.470 --> 00:06:59.550
things are very, very important in
these days. So that's also something to
96
00:06:59.589 --> 00:07:01.860
think about. When you're collecting data. Can you make it safe so your
97
00:07:01.939 --> 00:07:05.220
customers not doing any mistakes and just
coming to say hey, we just collected
98
00:07:05.300 --> 00:07:09.860
like DII and it's a bad news
for everybody. So and as well as
99
00:07:09.939 --> 00:07:12.699
security, as we said before,
E. Security is also an element when
100
00:07:12.699 --> 00:07:15.009
you're collecting data other people. We
want to get to that data and the
101
00:07:15.050 --> 00:07:19.009
way you collect data is also another
area that's hackers, for example, will
102
00:07:19.050 --> 00:07:23.769
find very interesting. So you want
to make sure that you aware of the
103
00:07:24.050 --> 00:07:29.009
security implication and make sure that you're
almost as as secured as possible and that
104
00:07:29.449 --> 00:07:32.040
respect. Yeah, definitely. I
mean secured is a big issue today,
105
00:07:32.120 --> 00:07:35.199
of course, and a lot of
challenges with privacy. On the side of
106
00:07:35.279 --> 00:07:39.319
scale, it's really interesting because we've
talked to a lot of companies and you
107
00:07:39.399 --> 00:07:42.680
know, a lot of companies struggle
with the scale and and reality of the
108
00:07:42.759 --> 00:07:45.350
World War and today, where there's
much more data than there was in the
109
00:07:45.430 --> 00:07:48.149
past. But the next layer they
were starting to hear a lot more about
110
00:07:48.350 --> 00:07:51.709
is the real time aspect. So
it's one thing to be able to store
111
00:07:51.709 --> 00:07:55.750
a lot of data and put it
a data warehouse somewhere. It's a whole
112
00:07:55.790 --> 00:07:59.339
other thing to be able to make
decisions based on data the streaming or a
113
00:07:59.579 --> 00:08:03.220
live environment. I'd love to hear
more about kind of how you've dealt with
114
00:08:03.379 --> 00:08:07.220
the the streaming or live data challenges. Yeah, absolutely. I think when
115
00:08:07.259 --> 00:08:11.860
you go into this the two paradigms. One is the big data analytics where
116
00:08:13.170 --> 00:08:18.449
customers expect to see data in seconds, minutes, sometimes hours. The aspiration
117
00:08:18.610 --> 00:08:22.250
is to be in the seconds really
space, I would say. But on
118
00:08:22.370 --> 00:08:26.449
the real time side, because as
a solution sometimes you want to trigger re
119
00:08:26.610 --> 00:08:30.920
time action. It might be just
you know, if you're optimizing your ads,
120
00:08:31.000 --> 00:08:33.000
for example, it might be that
you know you need to send the
121
00:08:33.080 --> 00:08:37.879
bid in the in re time.
Back then, when I dial with financial
122
00:08:37.919 --> 00:08:41.950
market, that's the real time decision
of buy and say, and today the
123
00:08:41.110 --> 00:08:45.830
real time might be just the right
campaign and want to serve for the users.
124
00:08:46.470 --> 00:08:50.110
And really time is in the in
the space, so on the the
125
00:08:50.190 --> 00:08:52.950
range of the millisecond real time.
And what you have to have in with
126
00:08:54.070 --> 00:08:58.620
real time is predictability, because you
have just a minimum window to take action
127
00:08:58.940 --> 00:09:05.059
and it has to be sustainable.
If suddenly one mistake there or one logical
128
00:09:05.100 --> 00:09:09.700
error, suddenly the dire real time
just become seconds. So for monitoring perspective,
129
00:09:09.740 --> 00:09:11.929
everything is green, no, but
nothing actually is not functioning. But
130
00:09:13.450 --> 00:09:18.169
from the business perspective, you're not
delivering the value. So as soon as
131
00:09:18.210 --> 00:09:22.210
you breached that milliseconds ranged into seconds, that is is a breach in your
132
00:09:22.330 --> 00:09:26.639
SLA. So there's many things to
think about when you think about real time
133
00:09:26.200 --> 00:09:31.200
action. How much data can you
analyze in real time? In theory,
134
00:09:31.279 --> 00:09:33.799
for example, if we have unlimited
time, we can take great decisions.
135
00:09:33.840 --> 00:09:37.480
We can scan ton of data in
real time. You can just cannot afford.
136
00:09:37.480 --> 00:09:41.309
You have to optimize for the best
decision within the time so you can
137
00:09:41.389 --> 00:09:45.350
close the little bit and do something
about that, versus, you know,
138
00:09:45.590 --> 00:09:48.990
just you know, spending a lot
of time but the customer is already gone.
139
00:09:48.110 --> 00:09:52.190
So your answer is great, but
there's not going to anyone to listen.
140
00:09:52.789 --> 00:09:54.980
So real time has a lot of
challenges and it starts from how you
141
00:09:56.179 --> 00:10:01.179
model the data, how you process
that, how you guarantee that response time
142
00:10:01.539 --> 00:10:05.820
SLA, for example, the predictability
in that, and how can you increase
143
00:10:05.860 --> 00:10:11.610
all the time the amount of data
you can trigger, you can actually analyze
144
00:10:11.730 --> 00:10:15.330
before you come back? And technology
is always evolving. So obviously using in
145
00:10:15.490 --> 00:10:20.289
memory is is giving us the that's
the fastest speed to analyze. But there's
146
00:10:20.289 --> 00:10:24.240
a lot to do with the way
you model in architecture solution to allow the
147
00:10:24.360 --> 00:10:28.440
best of the kind of allowing for
the best decision to be made within a
148
00:10:28.600 --> 00:10:31.279
very, very short time, and
they're very strict tessellary. So it's about
149
00:10:31.320 --> 00:10:35.590
the modeling the product, you know, the it's about optimizing the size of
150
00:10:35.710 --> 00:10:41.309
your data model, optimizing the execution
rules, you know, if we're,
151
00:10:41.389 --> 00:10:46.070
for example, running twenty, thirtyzero
rules per second in many, many cases,
152
00:10:46.789 --> 00:10:50.980
and and it obviously has to be
smart execution because eventually every user wants
153
00:10:52.019 --> 00:10:54.539
to get the right message of the
right time within the milliseconds, because are
154
00:10:54.539 --> 00:10:58.059
right now, are on the website
or in your product and you have just,
155
00:10:58.259 --> 00:11:01.179
you know, less than five hundred
milliseconds to close them. So that's,
156
00:11:01.299 --> 00:11:05.009
though, the the what we live
in and the fact that you need
157
00:11:05.090 --> 00:11:07.970
to think about the way you run
the you build a model. The run
158
00:11:07.090 --> 00:11:13.570
time is critical and many engineers just
are usually are kind of focusing on area
159
00:11:13.730 --> 00:11:18.049
when seconds is fine, you know, even three to five seconds everything.
160
00:11:18.129 --> 00:11:22.519
So it's to become an engineer that
knows how to deal with real time.
161
00:11:22.559 --> 00:11:24.840
You need to can shift the way
you think about everything you do, starting
162
00:11:24.879 --> 00:11:30.360
from coding, the size of object
you're building, the way you understand multi
163
00:11:30.480 --> 00:11:35.750
threading? All those elements we actually
learning university but scarcely apply in our day
164
00:11:35.789 --> 00:11:39.230
to day because most application are fine
with, you know, two or three
165
00:11:39.269 --> 00:11:43.870
seconds and your code can be very
inefficient and will still work, as opposed
166
00:11:43.909 --> 00:11:46.750
to real time when, if your
code is not in is inefficient, you're
167
00:11:46.750 --> 00:11:48.700
not going to be able to deliver
that at scale. Yeah, that makes
168
00:11:48.740 --> 00:11:52.779
sense. It sounds like what you're
describing a lot of times in the real
169
00:11:52.940 --> 00:11:58.139
time environment is really user experience.
So it's really the customization of how you're
170
00:11:58.179 --> 00:12:01.809
delivering to the user. And it's
one thing on a website with some content
171
00:12:01.970 --> 00:12:05.330
to serve up some better content right
past content, but it's a whole other
172
00:12:05.370 --> 00:12:09.289
thing when you're talking about a full
product experience. Right. Can you talk
173
00:12:09.330 --> 00:12:11.529
a little bit about about that and
where you see the evolution too, of
174
00:12:11.809 --> 00:12:16.080
where we're at, you know,
as far as what expectations are of the
175
00:12:16.159 --> 00:12:18.720
customer. So the expectation is again
to be able to do the the three
176
00:12:18.799 --> 00:12:22.440
things like one, to to understand
what the customer is because then it will
177
00:12:22.480 --> 00:12:26.759
drive the best decisions. If you're
limited in the what elements, what type
178
00:12:26.759 --> 00:12:31.470
of data input you have, then
the output won't be as effective. So
179
00:12:31.750 --> 00:12:35.309
number one is to really make sure
that we are collecting in real time as
180
00:12:35.389 --> 00:12:37.350
much data as possible. So we
want to understand what the customer is.
181
00:12:37.750 --> 00:12:41.350
Why did they do in the past
three months, for example, should so
182
00:12:41.549 --> 00:12:45.340
there's very simple use cases. I'm
building new features. How can I drive
183
00:12:45.379 --> 00:12:48.500
you to know those features and how
can I make sure that it's relevant for
184
00:12:48.620 --> 00:12:52.820
you? So I need to know
did you use it before, what type
185
00:12:52.820 --> 00:12:56.580
of usage pattern you have? All
that type is based on the decision,
186
00:12:56.740 --> 00:12:58.850
based on the collection of data that
I can do. It means that I
187
00:12:58.970 --> 00:13:03.610
want to understand, for example,
feature usage. I want to understand who
188
00:13:03.610 --> 00:13:07.090
you are as a persona this has
to be part of the model that you're
189
00:13:07.129 --> 00:13:09.970
building. So, for example,
in our solution we translate and call it
190
00:13:11.129 --> 00:13:13.440
every click they use a does to
a feature emodule. Also, we have
191
00:13:13.519 --> 00:13:16.720
better context of what you're doing right
now, because I don't want to recommend
192
00:13:16.759 --> 00:13:20.120
a feature you just clicked on.
You just went there and I'm like,
193
00:13:20.240 --> 00:13:22.919
Hey, learn about this, like
I already did that a second ago.
194
00:13:24.120 --> 00:13:28.830
So ast type of message, because
we know that today, as many ways
195
00:13:28.990 --> 00:13:33.710
to consume content, as many ways
to users choose different ways to use your
196
00:13:33.789 --> 00:13:37.269
product. So what would be the
best messaging? So the triggering of that
197
00:13:37.429 --> 00:13:43.059
message, it has to come to
come in into the to the product in
198
00:13:43.139 --> 00:13:46.259
real time, might be very different. So what you have to do is
199
00:13:46.340 --> 00:13:48.940
a top of the triggering rules,
you want to make sure that you have
200
00:13:50.419 --> 00:13:54.179
prioritization rules and you have some fadling
rules to make sure that you're not interfering
201
00:13:54.259 --> 00:13:58.769
with the user interaction, just helping
the user but not blocking the users.
202
00:13:58.850 --> 00:14:01.809
And if there's collision of different type
of messages, they might be all be
203
00:14:01.970 --> 00:14:05.409
relevant, but the expectation is the
system to be smart enough to choose the
204
00:14:05.490 --> 00:14:09.960
right one. So in some cases
we let you know the users prioritize,
205
00:14:11.039 --> 00:14:16.159
but the vision is to have the
system decide what's best, because you might
206
00:14:16.200 --> 00:14:20.200
qualify to more than one elements.
And the last thing is, I the
207
00:14:20.240 --> 00:14:22.720
expectation is to now that we are, you know, collecting so, so
208
00:14:22.799 --> 00:14:26.789
much data and a click of a
button today we can map your entire product
209
00:14:26.909 --> 00:14:31.710
and suddenly you can have visibility to
five hundred features. Were in the past
210
00:14:31.830 --> 00:14:35.429
you saw, for example, ten
features because you have your engineer collect and
211
00:14:35.549 --> 00:14:39.220
instrument ten features. But now,
suddenly, with automation, we can surface
212
00:14:39.220 --> 00:14:41.299
all your features, as a lot
of features to instrument. How can they
213
00:14:41.620 --> 00:14:45.779
make you smart about the data?
which features you should look at? What
214
00:14:45.899 --> 00:14:50.500
type of KPIS and patterns? Cause
as users we want to get the insights,
215
00:14:50.779 --> 00:14:52.610
not necessarily all the reporting. Right
you go. We don't want to
216
00:14:52.610 --> 00:14:56.850
have yet more data. We want
to be smart about the data. So
217
00:14:56.009 --> 00:15:01.009
Are our vision. Our goal is
to help you distill the right set of
218
00:15:01.090 --> 00:15:03.250
data so you can really be better
at your job. And that's it.
219
00:15:03.409 --> 00:15:07.840
That's a challenge because in data you
know that you need to have the right
220
00:15:07.919 --> 00:15:11.759
context and you need to know what
is the right thing to surface to a
221
00:15:11.919 --> 00:15:16.279
person to assist them with decisioning.
So, for example, what is the
222
00:15:16.399 --> 00:15:22.509
pattern of a good user behavior?
What is the pattern of a customer at
223
00:15:22.590 --> 00:15:26.309
risk? This type of things we
aspire to build using machine learning. Just
224
00:15:26.470 --> 00:15:30.190
understand how can I surface things to
you, because as human being, you
225
00:15:30.269 --> 00:15:33.870
cannot just digest so much data.
You have to get this assistance so I
226
00:15:33.950 --> 00:15:39.620
can point you towards a normally they
can point you towards like customers at most
227
00:15:39.700 --> 00:15:43.820
risk that will impact the most revenue
for you. Those type of decisions today
228
00:15:43.820 --> 00:15:48.580
are manual. Most companies are not
good collecting data. They usually collect what
229
00:15:48.700 --> 00:15:50.850
they can and then they search for
based on the data that they have,
230
00:15:52.090 --> 00:15:54.850
but they might be missing, you
know, around seventy percent of the potentially
231
00:15:56.009 --> 00:15:58.210
that they should look at. So
there's a lot of challenges and we are
232
00:15:58.250 --> 00:16:00.649
trying to kind of address all of
them, from collection to insights to drive
233
00:16:00.690 --> 00:16:06.159
action, and our vision is to
start to harness machine learning to help us
234
00:16:07.240 --> 00:16:11.480
be kind of drive better decisions for
our end users. Yeah, that's interesting.
235
00:16:11.559 --> 00:16:15.919
So how much are you relying on
machine learning models versus let's say,
236
00:16:15.000 --> 00:16:18.830
business rules or just simple regressions?
You Know How partant is that in your
237
00:16:18.870 --> 00:16:23.309
business? So today I would say
that it's still ninety percent like business rules,
238
00:16:23.389 --> 00:16:26.269
but more complex one as well,
and some of them are automatical,
239
00:16:26.509 --> 00:16:30.750
automatic and based on score, even
scoring. Today I think it's still eventually
240
00:16:30.750 --> 00:16:34.419
business rules, even if you build
very complex scoring, it's start to tat
241
00:16:34.539 --> 00:16:38.019
that. You were, you know, some of the kind of statistical modeling,
242
00:16:38.100 --> 00:16:41.659
but still it's pretty basic, cause
you are not the weights you're putting
243
00:16:41.700 --> 00:16:47.690
on, not necessarily the more gut
feeling or your experience versus they will data
244
00:16:47.730 --> 00:16:52.850
analysis that you should be doing.
Because so we basically want to do to
245
00:16:52.009 --> 00:16:56.889
give our customers the way to run
multiple models based on the data. What
246
00:16:57.009 --> 00:17:02.320
we sold for them is actually the
collection and the fact that we get clean
247
00:17:02.559 --> 00:17:06.559
data in if we are saying this
is a feature use, this is actually
248
00:17:06.559 --> 00:17:10.079
a feature. You don't need to
worry about the what we call the overfitting
249
00:17:10.160 --> 00:17:12.200
on too much, you know,
junk data. So the cleanness of the
250
00:17:12.240 --> 00:17:15.470
data is the one thing that really
helps us build better models. We do
251
00:17:15.670 --> 00:17:19.109
think that every company should have way
more than one model. When you building
252
00:17:19.150 --> 00:17:22.230
predictive models, the way I see
it, you need to slice the customer
253
00:17:22.309 --> 00:17:27.150
journey to, you know, the
different stages and build model to optimize for
254
00:17:27.269 --> 00:17:32.099
each. So if you try to
have any machine learning to solve a specific
255
00:17:32.220 --> 00:17:36.220
problem, I would say the you
probably going to have multiple models. And
256
00:17:36.339 --> 00:17:38.339
what we want to do is give
you the option to yes, tune the
257
00:17:38.420 --> 00:17:42.779
molds, run different molds and see
what actually performs best based on the goals
258
00:17:44.019 --> 00:17:47.490
that you have. For example,
if it's trial conversion and you're trying to
259
00:17:47.529 --> 00:17:48.970
build them all to say what would
be the best, you know, feature
260
00:17:49.009 --> 00:17:53.609
to promote or type of campaigns to
run, there's many model you can run.
261
00:17:53.809 --> 00:17:56.529
We are going to allow you to
run multiple and then you're going to
262
00:17:56.569 --> 00:18:00.720
see the result and you're going to
see their accuracy and decide what's best to
263
00:18:00.799 --> 00:18:04.400
you. And for onboarding, adoption, retention, all those areas are going
264
00:18:04.400 --> 00:18:08.920
to have different models because you might
be even different persona different segments of companies.
265
00:18:10.039 --> 00:18:11.240
One of the biggest mistake I've done, I've seen is, like you
266
00:18:11.319 --> 00:18:17.150
know, companies are building a holistic
adoption score one for all, and I
267
00:18:17.309 --> 00:18:19.789
think you need you need at least
twenty, because you have five person as,
268
00:18:19.789 --> 00:18:25.109
you have four four different cycles,
stages in the customer life cycle and
269
00:18:25.190 --> 00:18:30.619
you have did different inputs fir stage. So I think the what we're trying
270
00:18:30.660 --> 00:18:33.180
to do is to make that very
accessible for you to build models, run
271
00:18:33.259 --> 00:18:36.900
them, even if you're not the
data scientists, and, based on the
272
00:18:36.940 --> 00:18:40.740
accuracy, just decide what is the
best input and what is the best algorithm
273
00:18:40.779 --> 00:18:42.609
to use. That makes sense.
You know, with this, when you're
274
00:18:42.650 --> 00:18:45.250
when you're working with machine learning models, what it would you say is the
275
00:18:45.369 --> 00:18:48.730
number one challenge? You know,
sort of sort of working with AI machine
276
00:18:48.769 --> 00:18:52.450
learning. You know it's exploring everywhere, obviously, but there's still a lot
277
00:18:52.490 --> 00:18:56.119
of challenges in the road in the
roadway. You know what, would you
278
00:18:56.119 --> 00:19:00.640
say? Some of the significant ones
are absolutely I think if I look back
279
00:19:00.680 --> 00:19:04.480
almost would say a to ten years
ago when I first time, I stopped
280
00:19:04.519 --> 00:19:08.799
using machine learning. Back then I've
had some libraries, but nobody, you
281
00:19:08.880 --> 00:19:12.549
know, help you with collecting data. Nobody help you with running, for
282
00:19:12.630 --> 00:19:15.990
example, all the learning and testing. Basically you need to have this,
283
00:19:17.190 --> 00:19:21.309
the the whole process, to run
that. So you needed to build those
284
00:19:21.390 --> 00:19:26.019
data pipelines to go and basically having
the training set and testing set and so
285
00:19:26.539 --> 00:19:30.779
beyond even understand the whole concept.
It was just, you know, patching
286
00:19:30.819 --> 00:19:34.700
those solutions. And let's say you've
passed all those obstacles. Now you have
287
00:19:34.900 --> 00:19:41.329
a model that you want to stop
running data against it. What will help
288
00:19:41.369 --> 00:19:44.369
you serve that like just to make
it like? Cause, again, we
289
00:19:44.450 --> 00:19:48.930
were in the we built a recommendation
count recommendation engine. So for us is
290
00:19:48.130 --> 00:19:52.880
we need to ask the model and
get the response and render content based on
291
00:19:52.039 --> 00:19:56.680
that. So all these were you
had to build versus use any type of
292
00:19:56.759 --> 00:20:00.599
solution. And over time, we're
saying more and more problems are being sold
293
00:20:00.720 --> 00:20:03.440
and I think now in the last
couple of years, you're seeing how all
294
00:20:03.480 --> 00:20:07.829
the big vendors are now really helping
you run models at scale with a lot
295
00:20:07.829 --> 00:20:14.150
of data and even, in some
points, helping you serving and basically hitting
296
00:20:14.190 --> 00:20:18.829
these models and getting responses. So
a lot of obstacles are going down now.
297
00:20:18.990 --> 00:20:21.430
It's the time that we are think, you know, the the challenge
298
00:20:21.430 --> 00:20:22.859
would be like, how do you
choose the right data? How do you
299
00:20:22.940 --> 00:20:26.380
avoid overfitting? How do you avoid, you know, just being biased because
300
00:20:26.460 --> 00:20:30.500
you're bringing a model data set.
You're able to be to bring it not
301
00:20:30.579 --> 00:20:33.859
might not have the answer to begin
with. So common sense is always good
302
00:20:34.220 --> 00:20:37.170
and I've seen no, you know, Super Smart People. They just like
303
00:20:37.369 --> 00:20:41.609
the the the data source they use. They did not question the data source,
304
00:20:41.650 --> 00:20:44.769
they did not question their quality,
you know, they kind of looked
305
00:20:44.769 --> 00:20:45.769
at it as as kind of an
end game. This is the day that
306
00:20:45.769 --> 00:20:49.799
I have as opposed to know,
you should be exploring a lot of different
307
00:20:49.799 --> 00:20:53.079
data sets. Maybe the answer is
not there. That's a very common mistake.
308
00:20:53.160 --> 00:20:57.559
And second challenges again to understand and
make sense of it. And again
309
00:20:57.599 --> 00:21:02.720
sometimes you're getting the score and nobody
at instand why it is. So eitherly
310
00:21:02.720 --> 00:21:04.869
choose to trust it or not trust
it, and I think that if you
311
00:21:06.029 --> 00:21:10.549
break down two different scores, different
models, you'll be able to assess better
312
00:21:10.670 --> 00:21:12.990
that. This is why you're getting
this score as opposed to very high level
313
00:21:14.069 --> 00:21:18.430
kind of predictive score. And last
I would say it should be designed to
314
00:21:18.660 --> 00:21:22.779
assist your decision, not replace your
decision. Is always going to be a
315
00:21:22.819 --> 00:21:27.140
human factor that will be involved just
to understand and tune and feed the algorithm
316
00:21:27.140 --> 00:21:32.619
and decide to rerun the model.
By the way, there's another element that
317
00:21:32.779 --> 00:21:37.369
is very common in financial market.
There rebuilding the model almost nightly, and
318
00:21:37.529 --> 00:21:40.690
in the in the bit to be
industry, they build one model per year.
319
00:21:40.809 --> 00:21:44.170
It just you know, you basically
are. You have to understand that
320
00:21:44.289 --> 00:21:47.799
building models and very fine is somebody
that you need to do more frequently.
321
00:21:48.480 --> 00:21:51.400
Maybe not every twenty four hours,
but you know, at least waterly,
322
00:21:51.440 --> 00:21:53.640
at least monthly, and just morning
to the quality of the service you're getting
323
00:21:53.759 --> 00:21:56.519
from that. But I feel like
we are very getting, very, very
324
00:21:56.640 --> 00:22:00.400
close to that and the market is
ready to start understanding and using it,
325
00:22:00.430 --> 00:22:03.829
as opposed to the past that people
thought it's going to replace human beings.
326
00:22:03.869 --> 00:22:07.910
And I don't trust it and I
see a number and I can't make sense
327
00:22:07.990 --> 00:22:10.430
of it. So I think it's
going to be our future and you see
328
00:22:10.509 --> 00:22:12.950
more and more people are getting used
to assistance and we use you know,
329
00:22:14.349 --> 00:22:17.859
Google, Alexa and all those suff
that today's you know, the next generation
330
00:22:17.940 --> 00:22:23.259
is obviously completely trusting you know ai
technologies to to driving decisions without any question.
331
00:22:23.299 --> 00:22:26.940
Yeah, and I think Covid,
you know, changed a lot of
332
00:22:26.059 --> 00:22:29.500
things, obviously, and a lot
of ways. But I can tell you
333
00:22:29.660 --> 00:22:33.369
that with us and what we look
at in our customers, we saw a
334
00:22:33.529 --> 00:22:37.690
huge change in they're thinking. Or
because they experience Covid, they realize what
335
00:22:37.769 --> 00:22:41.009
you were just saying, which is, Oh, that model that we rebuild
336
00:22:41.049 --> 00:22:42.890
once a year, maybe we need
to do it more frequently, because they
337
00:22:42.890 --> 00:22:47.519
were totally at a loss, you
know, when covid hit, for what
338
00:22:47.680 --> 00:22:51.079
to do next or how to really
customize the experience for their customers. Right.
339
00:22:51.119 --> 00:22:55.480
Yeah, in many cases what happened
is like in most models, you
340
00:22:55.640 --> 00:23:00.190
basically looking backwards on data and driving
forward, and when covid hits, you're
341
00:23:00.190 --> 00:23:04.109
not going to see the tree understanding
that you need to refresh your model and
342
00:23:04.390 --> 00:23:08.789
move towards, you know, shorter
cycles and there's a whole element of new
343
00:23:08.950 --> 00:23:14.420
networks. But I think, but
I think in the classic data science you
344
00:23:14.740 --> 00:23:18.980
want to make shorter windows of time
to analyze the data so you're getting you're
345
00:23:18.059 --> 00:23:22.180
able to at least react. The
other thing is is now it's going to
346
00:23:22.220 --> 00:23:27.289
become much cheaper to use data science. I think it's more accessible. I
347
00:23:27.410 --> 00:23:33.089
think still I see companies that have
only say companies have the luxury and budget
348
00:23:33.170 --> 00:23:37.690
to have data science team and to
use, you know, those type of
349
00:23:37.970 --> 00:23:41.559
processes. But I think in the
future it's going to be available for for
350
00:23:41.720 --> 00:23:45.119
any size company to start utilizing machine
learning because because there's a lot of things
351
00:23:45.200 --> 00:23:48.279
to do. And the second thing
I see a lot is, again the
352
00:23:48.680 --> 00:23:52.079
scoring model. At Marquetto we had
at leade scoring, and most of the
353
00:23:52.079 --> 00:23:56.750
lead scoring I've seen by customers is
just that feeling of marketers. They say,
354
00:23:56.789 --> 00:23:59.630
if you clicked on this website,
you download this white paper, you
355
00:23:59.710 --> 00:24:03.549
qualifying like that's and even they might
enrich the data, but to really understand
356
00:24:03.670 --> 00:24:07.710
what behavior drives the best conversion,
you have to apply a machine and you
357
00:24:07.789 --> 00:24:12.460
have to apply much stronger dataalysis,
because you have millions of website visitors,
358
00:24:12.500 --> 00:24:17.099
for example, or millions of users, and there's no way you will be
359
00:24:17.180 --> 00:24:21.140
able to with your gut feeling to
score that. Today we're seeing with customer
360
00:24:21.220 --> 00:24:25.930
health score, we're with bullock adoption
scored as many scores that the data science
361
00:24:25.970 --> 00:24:30.609
teams are trying to do. So
when it's a data science team they tend
362
00:24:30.650 --> 00:24:33.450
to know exactly what they're doing in
the score tends to be pretty accurate.
363
00:24:33.849 --> 00:24:37.130
But not every company can afford a
full data science team and all that kind
364
00:24:37.130 --> 00:24:41.960
of data aggregation and processing, and
I think that's going to be something that's
365
00:24:41.960 --> 00:24:45.440
going to be more affordable to other
companies. Yeah, it's to playing a
366
00:24:45.519 --> 00:24:48.400
huge Doman for more data science talent, you know, as we move forward
367
00:24:48.440 --> 00:24:53.109
in the world hoping to make more
data driven decisions. Right. What advice
368
00:24:53.190 --> 00:24:56.990
would you give to someone seeking your
career in data science? You know,
369
00:24:56.069 --> 00:25:00.710
based on your experience, I think
that in data science and again, I
370
00:25:00.869 --> 00:25:04.910
always looked at data as kind of
what is the oil to drive business results.
371
00:25:06.109 --> 00:25:10.779
In marketing again, was the input
for personalization, for better statistics,
372
00:25:10.859 --> 00:25:14.579
for better insights about your customer intent. I think it's important for everyone,
373
00:25:14.660 --> 00:25:18.259
data science or engineering, just to
really understand and also care about the product
374
00:25:18.299 --> 00:25:22.170
and in the outcome. Just as
a general advice, but I think you
375
00:25:22.289 --> 00:25:26.089
know, it's becoming a very needed
and required talent in every company. So
376
00:25:26.369 --> 00:25:32.170
learning all the tools, applying this
different models and learning all the time is
377
00:25:32.210 --> 00:25:34.130
going to be very, very useful
because I think again, the market is
378
00:25:34.250 --> 00:25:38.880
now becoming ready for a machine learning
driven type of solutions, and I've seen
379
00:25:40.000 --> 00:25:42.960
those solutions actually around two thousand and
fourteen, like six seven years ago,
380
00:25:44.559 --> 00:25:48.680
there were a lot of machine learning
driven solutions, like lead predictively scoring in
381
00:25:48.799 --> 00:25:52.430
others that were way ahead of that
time and I think also they lack some
382
00:25:52.549 --> 00:25:56.950
of the technologies that exist today where
now they're very, very successful. But
383
00:25:56.069 --> 00:25:59.309
I think in two thousand and twenty
one, two thousand and twenty two,
384
00:25:59.430 --> 00:26:03.980
machine learning is becoming very except acceptable
and accessible and the need for data sciences
385
00:26:04.099 --> 00:26:11.140
to really be able to run those
type of motions are increasingly higher as well.
386
00:26:11.500 --> 00:26:15.339
I mean it's interesting that now,
as this whole industry revolves, what
387
00:26:15.460 --> 00:26:19.049
we're realizing is that this concept of
the growth mindset or learning or continuous learning,
388
00:26:21.250 --> 00:26:25.529
when you take people who are most
comfortable in math and technology and then
389
00:26:25.569 --> 00:26:29.049
you put them into an environment which
you have to or right to be in
390
00:26:29.089 --> 00:26:32.210
that world, but you put them
an environment where they're constantly have to having
391
00:26:32.250 --> 00:26:37.079
to evolve and change and and learn
about the industry and what's happening because,
392
00:26:37.079 --> 00:26:40.319
as you said, a few years
makes a huge difference, and what we
393
00:26:40.440 --> 00:26:42.599
can do and what the costs are
doing. That right, absolutely, and
394
00:26:42.640 --> 00:26:45.720
I think that they need two questions. A lot of the kind of facts
395
00:26:45.920 --> 00:26:49.549
to begin with, because again they
might just get a problem to solve and
396
00:26:49.589 --> 00:26:52.430
the data set, but they need
to question, is this really the data
397
00:26:52.509 --> 00:26:56.670
set? This is all we have? Can we actually exploded? We explore
398
00:26:56.670 --> 00:27:00.309
other things before you get to this
more optimization, because again, if with
399
00:27:00.549 --> 00:27:03.660
bed oil you can't really have a
great, you know, fast moving car.
400
00:27:03.900 --> 00:27:07.299
So it's that's that's something that you
want to make sure that you're always
401
00:27:07.299 --> 00:27:11.579
learning. And they need to also
understand the space, the problem that they're
402
00:27:11.660 --> 00:27:15.049
solving, because otherwise they would not
have these kind of curiosity questions. Again,
403
00:27:15.130 --> 00:27:19.130
I was in the marketing space or
in the financial space, understanding the
404
00:27:19.250 --> 00:27:23.809
paradigms there really helps you as data
sciences to understand this might be relevant data,
405
00:27:23.849 --> 00:27:27.690
this might be, you know,
less relevant data, and be more
406
00:27:27.730 --> 00:27:32.920
creative there, because eventually it requires
you to think what would be the best
407
00:27:32.920 --> 00:27:36.079
input to your models and why would
be the best models to fit. And
408
00:27:36.480 --> 00:27:40.880
again, without knowing the entire spectrum
of what you're solving, you might just
409
00:27:41.000 --> 00:27:44.390
looking under the fleshlight. Is supposed
to look where you know the data is
410
00:27:44.430 --> 00:27:48.750
really exists. Words representative. Yeah, that makes sense. Well, thank
411
00:27:48.789 --> 00:27:51.470
you so much for being on the
show today. You know, we really
412
00:27:51.509 --> 00:27:53.789
appreciate your time. Where can the
community find you? What's the best way
413
00:27:53.789 --> 00:27:57.269
to get in touch with you?
So I'm on Linkedin. My against it's
414
00:27:57.309 --> 00:28:03.019
Mickey alone on Linkedin at gain site
and my email is actually micky at gainsayecom.
415
00:28:03.140 --> 00:28:07.460
So feel free to reach out and
happy to advice. Also, we
416
00:28:07.660 --> 00:28:12.089
have a lot of content with regards
to analytic sport and it takes customer aalytics.
417
00:28:12.170 --> 00:28:15.650
So happy to if you have any
question around that, happy to take
418
00:28:15.690 --> 00:28:18.529
that as well. Okay, Great. Well, thank you. And at
419
00:28:18.569 --> 00:28:22.650
the data standard we are building a
community of data enthusiasts and data thought leaders.
420
00:28:22.690 --> 00:28:26.609
We look forward to engaging with you
on our website, data standard Ioh,
421
00:28:26.009 --> 00:28:30.599
where we regularly post a wide range
of podcasts with data executives and offer
422
00:28:30.680 --> 00:28:36.319
regular networking opportunities. For more information, visit the website. Thank you so
423
00:28:36.440 --> 00:28:37.680
much for joining us today. Thank
you for having me
