David Moghavem (00:17)
All right, welcome to another episode of DealFlow Friday. I'm your host, David Mogavum. Today we have Adam Rion, the founder and CEO of AKTUS AI. AKTUS AI is an incredible platform that we at Tryon are also using. They build custom AI solutions for real estate, private equity, any operators, workflows. they are an incredible team to work with, and Adam's been a pleasure to work with. So, Adam.
It's great to have you.
Adam Rian (00:48)
Hi David. Thank you so much for having us.
David Moghavem (00:51)
I wish we were able to do it in person, but with
Adam Rian (00:54)
Yes. Yeah.
David Moghavem (00:55)
scheduling conflicts, we're gonna do it virtually. But but it's
Adam Rian (00:58)
Next time we will do it in person.
David Moghavem (01:01)
I'm I'm excited for the audience to hear a little bit about AKTUS because we've engaged with you guys. It's been, I think we've been working together for about a month or two. Bruce Kirsch introduced us, who is one of the OGs of
Adam Rian (01:15)
Yes.
David Moghavem (01:16)
real estate with the
models and the tools that he's built. And it's actually interesting seeing how things are evolving. But so we'll get into that. but I wanted first for you to introduce yourself, introduce what AKTUS does, and we'll dive into it.
Adam Rian (01:32)
Absolutely. Thank you, David. So Actas is a startup in Silicon Valley started three years ago in order to build AI platform for financial companies, for heavily regulated industries, that the large language model by
By itself, which are mainly probabilistic, and we're gonna discuss about what does it mean and what are the downside, what's the good things about it, what's not necessarily good things about it, and how act us is gonna solve those specific problems. So that's AKTUS started three years ago to help regulated industries, specifically finance and and real estate, to use AI more effectively and actually.
finish the job, not just for conversation and answer and question. So actus is a Latin word for action, which is we are helping our client to take action and finish the work, not just having a conversation back and forth with AI. My background is mainly in AI and machine learning for the past 15 years. I got my PhD in in computer science, and and and helping research on
Publishing paper, understanding the foundation of natural language processing. And in the past 10 years, I was working in different industries, smaller and big. But now access is so exciting that I feel like we are helping our clients in a way that we can engage more and more specifically on regulated industries and real estate.
David Moghavem (03:09)
So I know part of your background, you were product lead for data and ML at Twitter, right?
Adam Rian (03:15)
Yes.
David Moghavem (03:16)
And part of the founding team and CTO at Flowcode. How has it been now transitioning to AKTUS actus where you're kind of working B2B?
Adam Rian (03:25)
Yeah. so in my previous roles, it was AI, but not necessarily B2B. One of the things that AI is changing, the B2B industry, is making these B2B softwares more exciting. Because historically, when you were building softwares and solutions for B2B, the people that they were signing a deal and they were buying your platform were different from the people that they were using the platform. Now
This is changing. The people who are purchasing the platform are gonna be the people who are gonna use the platform. So that's gonna make B2B platforms more like B2C. They need to really solve the problem. And the people who are purchasing and the people who are gonna make the decision, they're gonna use it themselves. So my background on the B2C now helping me to shape the product in a way that B2B a business.
And and the audience of the platform, they understand the value themselves, not just purchasing something for someone else to use it.
David Moghavem (04:35)
So let's go a little bit deeper into it. I mean, you tell me, correct me if I'm wrong, but a majority, if not all your clients are either in the financial or real estate sector. So
Adam Rian (04:45)
Yes.
David Moghavem (04:47)
why choose this industry?
Adam Rian (04:50)
so there are a couple of reasons. First of all, AI, out-of-the-box AI, cannot be helpful in this industry. So when you look at OpenAI and cloud out of the box, the way that they've built these systems, the way that they build this type of AI is based on the training of the data which was available openly to the public through internet.
So those types of large language models, they were able to read those data, understand it, and they perform action around those. But in heavily regulated industry, finance and real estate, most of the data are not public. And most of those decisions require some sort of a secret sauce, some sort of understanding, domain knowledge. And all of the these domain knowledge is going to be super crucial to make the decision.
And those things are not available. They were not available to the AI when AI were trained, and now it's not available to necessarily to the AI as well. So these types of custom solutions is gonna become more and more important for for for re specifically real estate. And the other thing is real estate is multidisciplinary industry. It's not just finance, it's finance, it's legal.
It's architectural decision sometimes, it's relationship-based. And we believe these types of multidisciplinary industries, including real
estate, they're gonna benefit from AI if we design it right for them. And that multidisciplinary nature of the real estate makes it complex for out-of-the-box AI solution to provide the just
simple solution, simple platform, and at the same time, there are lots of people in this industry that they need help. so it was kind of
We got pulled into real estate because as soon as we published our platform, we received lots of inbound from real estate investors that they want to use it. And then, as much as we are working with more and more real estate investors, we realize we can provide lots of help. So it was not from the design. We started AKTUS building AI platforms for heavily regulated industry. And through our working with different clients, we got pulled into real estate.
David Moghavem (07:19)
Right. I mean, that's the nature of being an entrepreneur and being part of the startup culture is you're continuing to adapt, you're continuing to find your niche in the market. One thing about AKTUS that I appreciate, and I've talked about this several times on the pod, the there is a moment with AI where you go from being just using AI to make everyday tasks more efficient to actually revolutionizing revolutionizing and redefining where.
You make your decision, data-driven decision making. I think with AKTUS, what's important, as you said, is it's a heavily regulated industry with the real estate. We all have our own proprietary data. To have Claude or one of these AI out-of-the-box solutions, as you mentioned, try to decipher through this type of data, it's not naturally positioned to process that data as well.
as maybe a custom AI solution to get deeper into coming up with investment theses, coming up with the right approach for a strategy. And so that's where AKTUS comes in, where they can actually tailor some AI solutions to a point solution for a problem that you may have.
Adam Rian (08:36)
Yes, yes, exactly. And the way that we are framing this, both for our internal team and engineers to understand and also for the customer is these type of AIs are your digital teammates. they are you are not purchasing new software or you are not purchasing a new application. You are you're building it, yes,
David Moghavem (08:55)
Right. You're building it. Yeah.
Adam Rian (08:59)
and you are training this as you are working with it. So these types of solutions.
If as you work with it more and more, they get more information based on the way that you are working with the system. And this the all of these AI, customized AI pipeline, they're gonna use that output, the way that you give it feedback, and then they go back and and retrain themselves. And we have to define the definition of training. And there are a couple of technical terminology regarding are you training the large language model itself or you are training the agent? We can get into this, but the
point for the business is these systems are getting better and better. The same way that when you are hiring an analyst, at the first month they're still trying to learn and they're trying to get better and better over time. But the way that from management perspective, you're gonna be happy about your hire is you see the progress. If you see random
Performance sometimes is good, sometimes it's not good, and then sometimes the performance of the financial model is not good, but sometimes it's good. That randomness, it's not good for the business, for the decision making. But even if at the beginning the performance is not good, but you see the progress, then it gives you the hope that yes, at some point it will kind of make some decision. But that progress needs to happen. And in order for that
Progress to happen, we should be able to measure it. And we have to design that measurement together. How, because that's different from one real estate investment firm to another real estate investment firm. They have to these they have to build their own specific metric, their own KPIs, and the AI solutions should follow those KPIs. And these are the the way that we can make sure we are building the right solution together.
David Moghavem (10:52)
I'm sure you get irritated when you hear the term vibe coding and someone's like, I could just vibe code this solution or I can just use Claw to do it. Where do you draw the line between what can really just be vibe coded to where it's like you need someone like actus to come in and solve that solution?
Adam Rian (11:10)
first of all, I love vibe
coding as a prototype. and vibe coding is a great way to communicate what people need. So before vibe coding, two years from now, when we were talking with a client, we need to sit down and have whiteboarding sessions and make sure we understand their workflow and actually what they need. And then based on those work sessions.
we were designing something inside Figma, and then we were presenting those in front of the client.
David Moghavem (11:43)
What's Figma? Sorry.
Adam Rian (11:45)
Figma is a design system, is a system
David Moghavem (11:47)
Okay.
Adam Rian (11:48)
that you can design a software or design a solution, but without writing any code.
David Moghavem (11:54)
Mm-hmm.
Adam Rian (11:55)
so that that so that was the thing that we were using before before vibe coding. and then after
couple of days of working with this type of design software, then we were presenting something in front of the client and the client would say, okay, yes, I I like this or not. But now they can vibe code things themselves. And in a first session with a client, we can sit down and we realize, okay, this is what they really need because they vibe coded in this specific way. So vibe coding is very, very helpful to understand the requirements.
And to understand what exactly can what type of product can solve their problem. But some people think it's like 80 20 rule that with vibe coding you can get to 80% of the platform, but it's gonna be very hard to finish that 20%. But I think we have to think about it from different perspective. So vibe coding is great for prototyping.
And requirement gathering and make sure we understand the pain and how that pain can get resolved. So vibe coding is 100% good for this one. But for actual production, it's 0%. Because we cannot use any code that happened inside vibe coding to the production code. We have to regenerate all of those codes.
David Moghavem (13:24)
Why is that? Why is it?
Adam Rian (13:27)
Because it's gonna be very hard to maintain. These types of softwares,
David Moghavem (13:31)
Mm-hmm.
Adam Rian (13:32)
a lot of things are happening on the back end. If you want to push it into the production, we have to make sure we can maintain it. And there are gonna be lots of edge cases happening when someone uses the platform. And in the AI, those and specifically in real estate, those edge cases are most of the
The work is gonna be around those edge cases. For example, when
David Moghavem (13:57)
Mm-hmm.
Adam Rian (13:58)
you are ingesting lots of lees, there are lots of things that we have to pay attention to those leaves, like appendix, like amendment, like something change, or some someone cross a number and add another number by hand. So all of those things the the platform should be able to ingest and and clean the data and present it in the right way. And white coding is gonna ignore all of those things.
But at the same time, from an engineering perspective, when we want to maintain a code, we should be able to understand it. And AIs, they tend to do lots of coding, even on a text, if you ask AI to write an essay, so it's gonna do lots of tokens, lots of words, lots of like 200 pages.
And we don't even have time to read all of those things. And sometimes you don't need those 200 pages, you just need maybe two pages to say what you want to say. The same thing will happen in vibe coding. When you vibe code something, and when we look at the source code, lots of those codes are unnecessary. And that creates future bugs. So we have to remove them. And in order to remove them, we have a process. So there is a harness engineering style of.
Coding that we are using internally, and lots of startups are using, and through that harness engineering, we are able to remove hallucination and make sure all of those AI-generated code can be used in production. So that's the process. We sit down with a client, we see what they did in a vibe coding session on their own, then we use that for the requirement gathering. Then it's gonna be a process to learn from
From product perspective, we're gonna learn from what has been five vibe coded and then we're gonna build it from a scratch in a production ready environment.
David Moghavem (15:54)
So I want to go back to the concept because I think it's important of how vibe coding can only get you so far with prototype, but the fact that you have to continually maintain it in order for it to be productive ends up being a detriment to your own productivity. And so that's
Adam Rian (16:14)
Exactly.
David Moghavem (16:14)
where actus can come in. That's where enterprise software can come in and actually make something that you vibe coded into a real sustainable platform.
That you don't need to be a real estate guy that's part-time engineer anymore. Like I've fallen into that trap before where I've maybe vibe coded a solution and then I spend more time trying to maintain it to make sure it works instead of that doing the actual production of what I'm trying to achieve. And so for those of you out there in our audience who feels like they're good with AI or strong with some of these out-of-the-box solutions.
I'm sure you've fallen into traps where you're like, why am I spending half my day trying to correct this to maybe be more efficient later, only to have the software, the code break on an edge case? so I think we're entering that phase where we're feeling the limits a bit on these out-of-the-box AI solutions. And if we want to actually take AI to the next step in our own businesses, we need to maybe.
let go of the the keys or let go of our own egos and see what's out there and almost level up with the code so that it's sustainable and can lead to real results.
Adam Rian (17:35)
That's correct, David. And I'm really happy that you you mentioned that. And we've seen lots of these products that they failed because of exact problem that you mentioned. it's give you lots of good feeling that I was able to achieve all of these things in just a weekend. but the real test is the productivity gain that you mentioned.
If
we're gonna if we want to get to that productivity gain, we have to design a system that can replicate your existing workflow. And we should be able to measure the success of that specific workflow. If we cannot measure it, we cannot improve it. And that's the crucial part. Measuring the success of this type of productivity tool is important, and that should happen from the design, not after the fact.
David Moghavem (18:31)
So going into measuring success success, and we actually have been going back and forth with your team about what defines true success with the platform. Because for instance, like, you know, we're doing a lead gen platform, we'll go into it a little more, but booking a meeting may be different than booking the right meeting, right? And so how do you guys approach defining customer success with your guys' product?
Adam Rian (19:01)
so that's a great question, and that's gonna be through onboarding process. when we are talking with real estate firms, so there are a couple of important things that they are paying super attention to. One is the when they do the deal screening, they want they don't want to waste their time on a deal that they should not spend any time on.
So they want to minimize the amount of time that they're spending on it, not a good deal. And that can be solved with the right AI tool. Then we need to measure, like some people, they are so good and understanding is it a good deal or not, or should I go deeper? Should I ask for data room? Should I sign an NDA or not? Then without AI is just a couple of seconds for them. But some people they need to spend maybe half an hour.
And do a little bit more analysis on that. So these are different from one person to another person, from one firm to another firm. And we want to make sure this new AI workflow respects what they are actually doing right now. Then at the same time, that one person, which is very, very good, their judgment is amazing on understanding is the good deal or is a bad deal, how we can translate that judgment into real AI workflow.
When that specific person they have an intern, what are the training mechanisms between the senior person and a junior person? That training mechanism we have to apply it into this type of AI workflow. Then we have to actually measure the time, how long it's gonna take to successfully screen a deal or successfully underwrite a deal, and how long it's gonna take to
Make sure we are writing down a right IC memo. so these times are very important. But the real metric is if that deal was really good or there is something was hidden that we didn't know about it. And sometimes AI makes mistakes, the same way that humans does make mistakes. And since we are early on and building this type of AIs, sometimes
We have to wait and let's say a year from now or two years from now, go back and evaluate all of those decisions that AI made. And that's easier when you have digital teammates or AI workers compared to the human. Because then when we have analysts, there are lots of things that might change. People leave, people you you hire a new people, lots of institutional knowledge will leave the organization when people are leaving.
But when you build this type of AI system, everything has been logged, and every you can trace back every single decision to a specific judgment that that AI has been trained based on. And because all of things have been logged, then we can go back two years ago and ask a question why we made this decision and what we learned from it. So all of those things we need to design it from scratch before start building this type of AI agent.
So time to make a decision is one of them, but at the end, the quality of those decisions are important.
David Moghavem (22:23)
Right. And the only way to know the quality of those decisions is the iterate reiterative process of going back to previous judgments, seeing what the outcomes are, and continuing that iteration.
Adam Rian (22:36)
Exactly. And how we can build a system that AI can learn from those iterations. So th so these are the things that we have to build when we are thinking about a digital teammate.
David Moghavem (22:48)
Right. It's it's a new way of benchmarking, right? As you said before AI, when a company would reflect ways for themselves to improve their own customer success from a human capability, there were so many variables changing. It could be a employer or employee that's coming in and out, or it could have been a bad employee or an outstanding employee that would sway whatever the metrics were. But now that everything is digitized.
and logged you can use that information to continue refining and re reinventing the way that maybe a workflow can be done.
Adam Rian (23:27)
Exactly. And one example is like financial models. So you mentioned Bruce Kirsch from our team. He's one of the top persons building this type of financial model and teaching the financial model. And we are so fortunate to have him in our team. So we spend lots of time together in order to see how we can design an AI system that can evaluate a financial models. What's a good financial model versus what's not optimum financial model?
How we can improve that. But what is a financial model at score? It's an abstraction layer that's gonna abstract all of these complexities that real estate has and then abstract it to verifiable numbers that we can use those in order to make that decision.
So someone has in basic couple of I think maybe seven years ago or even more, they invented a discounted cash flow model. So that's something that the mathematicians think through and invite in order to help people to make those decisions faster. So these are the abstraction layer that's gonna help us to make those decisions faster. So number one thing for the AI workflow.
is make sure it understands all of those existing abstraction layers and work with them. When you have a financial model, which is pretty complex financial model, you have to make sure AI understands it and AI can work with it. Then at some point you're expecting AI to give you hints and give you recommendation to improve your financial model.
And that's that's the point that we are designing the system with the client for the AI to get enough feedback to provide a guidance to change the financial model. We are still not at the stage that led the AI to make those edits by itself. We are at the stage that AI will give the recommendation and a senior analyst make those edits.
David Moghavem (25:30)
When do you think we'll ever get to that next stage? If if ever.
Adam Rian (25:35)
I think we will get there in less than five years. And the reason I'm using that five years is we need enough data. We need enough, and and this this is not universal. Some firms they might get there faster, some firms they might get there later. What the things that's gonna determine how fast we can get there is the high-quality data. So we need to give AI the chance.
To make a decision and learn from it. But humans should override those decisions right now. But we can do retrospectively and say, if we made that decision, what would have happened? And all of those things need to get logged because retrospective analysis is very important when you are using this type of AI. And then this type of retrospective analysis will provide more data for AI to learn. So when
real estate firms, they start using this type of AI agent now, I believe less than five years it will get to the point that it can make significantly good decisions. Better than
David Moghavem (26:45)
'Cause there's enough runway
to make retrospective analysis
Adam Rian (26:49)
exactly.
David Moghavem (26:49)
by having that window of of a five year period. Right.
Adam Rian (26:52)
Exactly, yes.
David Moghavem (26:53)
Yeah. It's it's gonna be an interesting time where we almost give the keys to AI to make those decisions and almost trust AI more than we trust our own gut or feeling. And it's some will say we'll never get there, some will say we'll get there soon. And but I do agree that that window
is a good benchmark for saying we'll have the data now where the retrospective analysis will become so strong and the iterative process will have occurred so many times. Just
Adam Rian (27:23)
Yes.
David Moghavem (27:24)
like you have a entry level analyst who's been working at the company for a long time, you don't need to correct that analyst as much or ever when they do that
Adam Rian (27:33)
Yes.
David Moghavem (27:34)
task for the hundredth time versus the first time.
Adam Rian (27:37)
Yes. I want to clarify something. Trusting AI more than trusting ourselves is a very broad term. but if we yes, if
David Moghavem (27:46)
Right. It's loaded. It's a loaded term. Yes.
Adam Rian (27:49)
but if we change the scope of those type of decisions and and then we can make decisions and this type of prediction better and faster, like in a financial modeling.
for real estate screening. So if we narrow down this scope, then it's gonna be easier for us to look at the data and see what that decision means and how we can teach AI to make those types of decisions. I believe human will make decisions much, much faster with these AI tools even now. And as we are building this type of system, then human
We'll make other types of decisions that we still don't know what those decisions are. Because our job descriptions are changing. Even inside
David Moghavem (28:39)
Mm-hmm.
Adam Rian (28:40)
our firm, when we are hiring a new person, their job description is different from 12 months ago. So and those
David Moghavem (28:46)
That's interesting. That's interesting.
Adam Rian (28:48)
those this there will be some decisions that human is gonna make in future that's gonna be significantly better than AI, but we still don't know what those decisions are.
David Moghavem (29:00)
Yeah, it's almost like we are there's going to be decisions in the future that we're making that we don't can't even fathom making them today. Mm-hmm. Could
Adam Rian (29:08)
Today, ex exactly. And those
require those abstraction lay.
David Moghavem (29:13)
Right. Could be algorithmic decisions and just like seventy years ago before Excel was born, decisions were made completely different than seventy years later, where we're deep into Excel models and we're making decisions on assumptions and growth projections that maybe in a few more years from now we're gonna be making decisions that are don't exist today.
Adam Rian (29:38)
Exactly, exactly. Like the like the question is, can AI invent something better than Excel? Because
David Moghavem (29:46)
Mm-hmm.
Adam Rian (29:47)
we as a human, we invented Excel, which is an amazing tool. Now we have an AI which we are training right now inside Actus that can operate inside Excel, which is amazing. So AI that can operate inside Excel and do all of those calculations and help you make a decision, this is an amazing achievement.
But the next level of achievement is AI invents something better than Excel, which hasn't happened yet, but it might happen in the future.
David Moghavem (30:18)
Right. I want to get back to the practical use cases that AKTUS has been able to accomplish with their clients. Maybe give an example of one of the projects you guys have worked on that you in A enjoyed doing, but B saw a tremendous amount of success that you've achieved for the client by doing so that they wouldn't have otherwise achieved if they didn't use you guys.
Adam Rian (30:45)
yes, I can talk about a couple of examples. most of them are around document operation in
David Moghavem (30:52)
Mm-hmm.
Adam Rian (30:53)
in in real estate and investment management. So in real estate.
When you want to make a decision, there are thousands of the documents that needs to needs to be read, understood, abstracted, extract some data from, and then make a decision. And those documents are large and huge. And not all of them are financial documents. Some of them are legal documents, like zoning, the zoning legal documents, lease abstract, lease documents, and also the financial documents.
documents. And right now, when you drag
and drop all of those documents in out of the box AI model, they will get hallucinated. And we can talk about a technical limitation which is around the context window that we can talk about it later.
David Moghavem (31:50)
We
we could we could bring it up maybe like a just for the audience who doesn't know like simple, short answer, like why do they hallucinate, you know?
Adam Rian (31:59)
So all of these AI, the same as our brain, they have a memory limit. if you give them more data that you can fit into their memory, they can start hallucinating. they call that a context window. How much what's the amount of data that you can give to AI and what's the amount of data that AI can give back to you? So that's that's the context window. And
David Moghavem (32:26)
It's almost
like if you're using one chat on Claude to do all your tasks, then it's just gonna hallucinate or get it wrong and that's yeah.
Adam Rian (32:35)
Exactly, exactly.
When when you create a new session, you reset the the context. So the more context means you have to pay more because you have to pay more token to to to AI providers. So both from performance perspective and cost perspective, we have to do context management. there is another part of this type of AI that although they are claiming
their context window can be up to one million token, which kind of equivalents to the keywords, but only 10% of that context window is effective.
David Moghavem (33:20)
That reminds
you that reminds me of the ice stat that your brain you only use ten or twenty percent of the capability or capacity of your brain. Yeah.
Adam Rian (33:28)
That's similar. Yes, that's
that that's something similar. So when you are reading a novel and then a week later you don't remember everything. And most of the time you forget things that happen at the middle. Most of the time you remember a beginning of the novel and end of the novel. And interestingly, AI has this the same type of behavior. When you give it lots of data, AI tends to forget things that it's at the middle of it.
David Moghavem (33:55)
The similarities between like the human brain and the way AI h remembers things. It it's incredible. It's it's yeah.
Adam Rian (34:03)
It's fascinating. Exactly.
It's fascinating. so now the same way that when we hire someone or when we are educating someone, we try to give them the right amount of data at the right time, not everything at once, the same techniques that we have to use for AI.
And that's what Act as AI is doing with a client. So when we are onboarding a client, we make sure we build and customize a context engineering platform for their need to make sure AI sees the sees data that they they need to see and when they need to see.
And it's sometimes it's very important to understand there's some part of the data that is better for you to hide from AI, because otherwise it will hallucinate. When you're talking about the finance, what part of the legal document that you have to show to the AI? So these types of things are are make the AI system more robust. And one of the the sucks success projects that we had in the past.
Was related to real estate investment decision, which we ingest legal document like zoning, lease document, and a financial document, and thousands and thousands of those documents. So on a back end, we were able to build an engineering system that slice and dice data and make a decision. The AI, which in this example is large language models, an LLM, like Cloud or OpenAI, they see what data and when to see it.
Then
they provide an answer. One of those input data to the system was a real estate financial model with 60 plus tabs, pretty complex comprehensive underwriting model. That was another input to the system. So right now, if you if you upload a comprehensive financial model into the cloud or OpenAI, they they hallucinate a lot.
Complex to understand that financial model. We have to build an engineering system to break those problems in the smaller piece and make sure AI understands it, then test it to make sure AI passed those tests, then use it in the underwriting process. So that was a summary of one of the success projects that we had in the past.
David Moghavem (36:35)
Right. I think it's important understanding you you went deep into a talk context management, where you can't just that's a slang word I've heard is one shot. You can't just one shot for one of these out-of-the-box AIs to do everything for you. It's gonna hallucinate, it's gonna make errors. It feels like it has to answer. It's almost like and it answers with confidence, right?
Adam Rian (37:01)
Yes.
David Moghavem (37:02)
And so you don't.
It doesn't say I'm not sure, but I think it's this answer. It answers in confidence because that's just the way it's been trained to, and that leads to a lot of errors. And that's where you guys can come in and point out the holes in AI and the fallacies and be able to optimize for it to be right and certain.
Adam Rian (37:25)
Exactly. So AI is an amazing tool. We need to use it in a right way. And that right way is this context engineering platform. And we call it when you harness OpenAI, when you harness cloud, then it's it's it's it's gonna solve those those those problems for you. There is no magic, there is no
one-shot learning. It's always requires back and forth. Human is not gonna happen for human it doesn't happen one shot. For AI is not gonna happen one shot. And these type of decisions in f in in investment is pretty complex. We have to take our time and make sure AI get it and and and we test it before we give it the specific tasks.
David Moghavem (38:07)
Right. So one thing we're using with you guys is we're building a lead generation platform for third sourcing third party management. And sourcing can be for third party management, can lead to acquisitions, loan sourcing. But the idea is that the way in which we're going about it is not your simple email saying, Hey, are you looking for management? Here's try on living. It's something that's a little deeper. It's actually ranking the properties.
whether it's close to our own portfolio that is performing well or other properties in different markets. And we're using as an anchor point to source and power rank the properties from you know best to worst, having different approaches for all of them and being able to have AKTUS actually thoughtfully reach out to prospective clients and see where we can actually add value, not just do a spray and pray.
And you can only imagine how trying to vibe code that is a disaster. You can maybe vibe code a component of it, maybe a type of deliverable, or maybe you can pull a certain contact, but to do it at scale is just a whole different project. And I think that's where for us we saw so much value with you guys that trying to vibe code that would be a disaster and having you guys on board has been has been great. So I'm excited to see where.
Where that goes.
Adam Rian (39:37)
Absolutely.
Thank you, David. And that's a great example of how digital teammates work. Let's say if you want to hire an analyst to do that task, then all of those things that you mentioned right now, you will let them know during the onboarding or even during the interview that these are the things that should happen. We should not spray and pray. We have to customize it. These are the ways that we have to do the search. So all of those things that you just mentioned.
These are the learnings and guardrails for AI. Now we're gonna let AI to make some decision, to do the sorting, to draft an email. So, as you know, we are not gonna let AI to send email yet. We're gonna let AI to draft the email. And those the draft, the input to those drafts is all of those judgments and knowledge.
And all of those documents that need to be ingested, and the AI should read them, understand them, the financial, the location, the zoning, and make sure that's the right prospect. And then, based on all of those, write down a customized email. Like, hey, I know you are doing specific things. We are your neighbor in that specific area. We did all of those things. And all of these emails are gonna be customized and changed.
Based on the prospect. And we are not at the point that's gonna let AI to send all of those things. The AI is gonna just draft them and make sure someone like you is gonna read them and review them. And then based on those feedback, you might say, okay, 90% of them are amazing. I'm gonna hit send, but the 10%, the AI should go back and redo it.
David Moghavem (41:20)
Right. And I think it saves the time. So now when you're sending that email, let's say the person doesn't respond, deletes the email, which is happening all the time.
Adam Rian (41:29)
Yes.
David Moghavem (41:30)
Rather than just leaving it at that, now you have a little bit of context of yourself of who that owner is, where they've transacted. You can pick up the phone, you could call, you can leverage some of your own connections. So it saves so much time from doing the lead cracking and the email writing.
and the and the sourcing to now you have that time and you can dedicate it towards finding the right relationship, talking to the right people, leveraging your own connections. And I think that's the next step for humans is as you said, our decision making and our lives are going to look a lot different in the next couple of years to where it was just a few years before. We're going to be doing less of this type of robot AI work and we're going to be doing more
Adam Rian (42:16)
Yes.
David Moghavem (42:17)
of this
Human leveraging work.
Adam Rian (42:19)
Yes.
Yes, yeah. And some people might think, okay, we are building a new CRM. No, that's not the case. It's it's a digital teammate. It's like
David Moghavem (42:29)
Right.
Adam Rian (42:29)
someone who are going to execute those type of tasks under your guidance. Sometimes those tasks are evaluate this OM, run this financial model on that data room, and give me the report. Sometimes the task is find these 20 prospects in this specific area.
So these are the tasks that our AI system is gonna help you to execute, not a new tool, not a new software, not a new application. And that's the difference that I want to make sure our audience understand this is a new paradigm of executing tasks. And in near future, I believe these type of AI agents, they demand their own application.
So all of these applications that we had in the past, they've been designed for humans to operate. Now we are building digital teammates together. And those digital teammates, they demand a specific type of tools. And those tools, we don't know what those tools are yet, but we have to work together in order to understand those. And then maybe sometimes we have to go and purchase a specific tool for our AI agents.
David Moghavem (43:45)
Yeah, I mean, like one thing that for me that rings the bell is like API keys, right. Where a human wouldn't need an API for their own. They could just go on to the UI and look at the data. But I think for efficiency's sake, having the right API keys and integration, I think that's the next step. I was actually saying how some of the tools we use that don't use or share their API key, I think they're behind. They're gonna end up getting lost in the sauce with what's going on.
where if you're leveraging AI, a tool that doesn't share their their API or share their tools almost can be become obsolete.
Adam Rian (44:25)
Exactly, exactly. So in future, we are going to have all of these digital teammates communicating with each other and working with each other. The the people that embrace this type of new transaction will win the market. And we have to design the system, we have to design a software that embrace a communication between different AI agents.
David Moghavem (44:49)
I wanna wrap it up with a question that I wanted to ask you. What's one thing about AI and real estate that's overhyped and one thing that's maybe underhyped?
Adam Rian (45:01)
So I would say most of the thing, most of the new technologies are overhyped in short term and underhyped in long term. So if we use internet for short term, for that specific couple of years, it was overhyped, but long term it was underhyped. So I believe specifically for for real estate, just drag and drop and one shot dealer screening is overhyped.
This is not something that people can rely on in a long term. Lots of people are talking about it, but I don't think it's gonna work in in in in in real world. But in long term, even that's gonna be under hype. and by long term, I think we have to wait maybe 12 months, 12 to 24 months for a couple of those.
overhyped, like out of the box one shot that you mentioned just one shot dealer screening. So these type of things will go away and customized solution for dealer screening, AI powered customized solution for dealer screening and asset management will replace those type of one shots. But overall I believe AI, not just for real estate, as a whole, as a new technology, is overhyped in short term, is underhyped in long
David Moghavem (46:24)
Well said. I think I agree with that. and so it'll be interesting to see where things evolve. It'll be very interesting to see how AKTUS evolves over over the next, you know, years and decades and and this in this cycle. Where what do you see next, I guess, for actus in these next twelve to twenty-four months as we go from a overhype to underhype type of environment with AI.
Adam Rian (46:51)
We
believe our customer will design their new org charts based on Act as digital teammates. So one of the things that AI will change the industry is the org chart itself. So why we have these specific org charts right now, because of our education system. We went to a school, we got architectural degree or finance degree or legal degree, and based on that.
The whole enterprise has been designed around those specific skill sets that we learned through school. AI is changing the whole education industry. And multidisciplinary's task is gonna be more achievable.
In near future. And those multidisciplinaries is gonna change the education system, is gonna change the skill set of the people. And in near future, we have to rethink how we are designing our orch chart. Do we need a marketing department separate from finance department? Or those orgs will be around actions, will be around the task or job to be done, as opposed to a skill set of humans. Our vision.
Our roadmap for Actas is in near future, is gonna influence the org chart of of enterprise and and they are gonna use Act as digital teammate to do the operation.
David Moghavem (48:14)
That's that's a a world where it's gonna be completely different, right? Digital teammates, having you know, basically AIs. I mean, I don't wanna say replace because as you said, I do think the human is gonna just also evolve to what they're gonna be doing, but it is replacing the jobs today. But I think it's also creating the jobs for tomorrow. So it's
Adam Rian (48:39)
Exactly.
David Moghavem (48:40)
gonna be interesting to see how AKTUS plays a role in that. Adam.
Really great having you on. Loved how we got theoretical with AI and practical with AI in real estate. for those of you who couldn't tell, Adam is also Persian brother. So Adam, Khali Khoshumadid,
Adam Rian (48:59)
Mm-hmm.
David Moghavem (49:00)
great having you on and looking forward to linking up soon in person and continuing working with the actus team on our project.
Adam Rian (49:08)
Absolutely. Thank you so much, David, for having me.
David Moghavem (49:10)
Awesome. Thanks.
Adam Rian (49:12)
Thank you.
All right, welcome to another episode of DealFlow Friday. I'm your host, David Mogavum. Today we have Adam Rion, the founder and CEO of AKTUS AI. AKTUS AI is an incredible platform that we at Tryon are also using. They build custom AI solutions for real estate, private equity, any operators, workflows. they are an incredible team to work with, and Adam's been a pleasure to work with. So, Adam.
It's great to have you.
Adam Rian (00:48)
Hi David. Thank you so much for having us.
David Moghavem (00:51)
I wish we were able to do it in person, but with
Adam Rian (00:54)
Yes. Yeah.
David Moghavem (00:55)
scheduling conflicts, we're gonna do it virtually. But but it's
Adam Rian (00:58)
Next time we will do it in person.
David Moghavem (01:01)
I'm I'm excited for the audience to hear a little bit about AKTUS because we've engaged with you guys. It's been, I think we've been working together for about a month or two. Bruce Kirsch introduced us, who is one of the OGs of
Adam Rian (01:15)
Yes.
David Moghavem (01:16)
real estate with the
models and the tools that he's built. And it's actually interesting seeing how things are evolving. But so we'll get into that. but I wanted first for you to introduce yourself, introduce what AKTUS does, and we'll dive into it.
Adam Rian (01:32)
Absolutely. Thank you, David. So Actas is a startup in Silicon Valley started three years ago in order to build AI platform for financial companies, for heavily regulated industries, that the large language model by
By itself, which are mainly probabilistic, and we're gonna discuss about what does it mean and what are the downside, what's the good things about it, what's not necessarily good things about it, and how act us is gonna solve those specific problems. So that's AKTUS started three years ago to help regulated industries, specifically finance and and real estate, to use AI more effectively and actually.
finish the job, not just for conversation and answer and question. So actus is a Latin word for action, which is we are helping our client to take action and finish the work, not just having a conversation back and forth with AI. My background is mainly in AI and machine learning for the past 15 years. I got my PhD in in computer science, and and and helping research on
Publishing paper, understanding the foundation of natural language processing. And in the past 10 years, I was working in different industries, smaller and big. But now access is so exciting that I feel like we are helping our clients in a way that we can engage more and more specifically on regulated industries and real estate.
David Moghavem (03:09)
So I know part of your background, you were product lead for data and ML at Twitter, right?
Adam Rian (03:15)
Yes.
David Moghavem (03:16)
And part of the founding team and CTO at Flowcode. How has it been now transitioning to AKTUS actus where you're kind of working B2B?
Adam Rian (03:25)
Yeah. so in my previous roles, it was AI, but not necessarily B2B. One of the things that AI is changing, the B2B industry, is making these B2B softwares more exciting. Because historically, when you were building softwares and solutions for B2B, the people that they were signing a deal and they were buying your platform were different from the people that they were using the platform. Now
This is changing. The people who are purchasing the platform are gonna be the people who are gonna use the platform. So that's gonna make B2B platforms more like B2C. They need to really solve the problem. And the people who are purchasing and the people who are gonna make the decision, they're gonna use it themselves. So my background on the B2C now helping me to shape the product in a way that B2B a business.
And and the audience of the platform, they understand the value themselves, not just purchasing something for someone else to use it.
David Moghavem (04:35)
So let's go a little bit deeper into it. I mean, you tell me, correct me if I'm wrong, but a majority, if not all your clients are either in the financial or real estate sector. So
Adam Rian (04:45)
Yes.
David Moghavem (04:47)
why choose this industry?
Adam Rian (04:50)
so there are a couple of reasons. First of all, AI, out-of-the-box AI, cannot be helpful in this industry. So when you look at OpenAI and cloud out of the box, the way that they've built these systems, the way that they build this type of AI is based on the training of the data which was available openly to the public through internet.
So those types of large language models, they were able to read those data, understand it, and they perform action around those. But in heavily regulated industry, finance and real estate, most of the data are not public. And most of those decisions require some sort of a secret sauce, some sort of understanding, domain knowledge. And all of the these domain knowledge is going to be super crucial to make the decision.
And those things are not available. They were not available to the AI when AI were trained, and now it's not available to necessarily to the AI as well. So these types of custom solutions is gonna become more and more important for for for re specifically real estate. And the other thing is real estate is multidisciplinary industry. It's not just finance, it's finance, it's legal.
It's architectural decision sometimes, it's relationship-based. And we believe these types of multidisciplinary industries, including real
estate, they're gonna benefit from AI if we design it right for them. And that multidisciplinary nature of the real estate makes it complex for out-of-the-box AI solution to provide the just
simple solution, simple platform, and at the same time, there are lots of people in this industry that they need help. so it was kind of
We got pulled into real estate because as soon as we published our platform, we received lots of inbound from real estate investors that they want to use it. And then, as much as we are working with more and more real estate investors, we realize we can provide lots of help. So it was not from the design. We started AKTUS building AI platforms for heavily regulated industry. And through our working with different clients, we got pulled into real estate.
David Moghavem (07:19)
Right. I mean, that's the nature of being an entrepreneur and being part of the startup culture is you're continuing to adapt, you're continuing to find your niche in the market. One thing about AKTUS that I appreciate, and I've talked about this several times on the pod, the there is a moment with AI where you go from being just using AI to make everyday tasks more efficient to actually revolutionizing revolutionizing and redefining where.
You make your decision, data-driven decision making. I think with AKTUS, what's important, as you said, is it's a heavily regulated industry with the real estate. We all have our own proprietary data. To have Claude or one of these AI out-of-the-box solutions, as you mentioned, try to decipher through this type of data, it's not naturally positioned to process that data as well.
as maybe a custom AI solution to get deeper into coming up with investment theses, coming up with the right approach for a strategy. And so that's where AKTUS comes in, where they can actually tailor some AI solutions to a point solution for a problem that you may have.
Adam Rian (08:36)
Yes, yes, exactly. And the way that we are framing this, both for our internal team and engineers to understand and also for the customer is these type of AIs are your digital teammates. they are you are not purchasing new software or you are not purchasing a new application. You are you're building it, yes,
David Moghavem (08:55)
Right. You're building it. Yeah.
Adam Rian (08:59)
and you are training this as you are working with it. So these types of solutions.
If as you work with it more and more, they get more information based on the way that you are working with the system. And this the all of these AI, customized AI pipeline, they're gonna use that output, the way that you give it feedback, and then they go back and and retrain themselves. And we have to define the definition of training. And there are a couple of technical terminology regarding are you training the large language model itself or you are training the agent? We can get into this, but the
point for the business is these systems are getting better and better. The same way that when you are hiring an analyst, at the first month they're still trying to learn and they're trying to get better and better over time. But the way that from management perspective, you're gonna be happy about your hire is you see the progress. If you see random
Performance sometimes is good, sometimes it's not good, and then sometimes the performance of the financial model is not good, but sometimes it's good. That randomness, it's not good for the business, for the decision making. But even if at the beginning the performance is not good, but you see the progress, then it gives you the hope that yes, at some point it will kind of make some decision. But that progress needs to happen. And in order for that
Progress to happen, we should be able to measure it. And we have to design that measurement together. How, because that's different from one real estate investment firm to another real estate investment firm. They have to these they have to build their own specific metric, their own KPIs, and the AI solutions should follow those KPIs. And these are the the way that we can make sure we are building the right solution together.
David Moghavem (10:52)
I'm sure you get irritated when you hear the term vibe coding and someone's like, I could just vibe code this solution or I can just use Claw to do it. Where do you draw the line between what can really just be vibe coded to where it's like you need someone like actus to come in and solve that solution?
Adam Rian (11:10)
first of all, I love vibe
coding as a prototype. and vibe coding is a great way to communicate what people need. So before vibe coding, two years from now, when we were talking with a client, we need to sit down and have whiteboarding sessions and make sure we understand their workflow and actually what they need. And then based on those work sessions.
we were designing something inside Figma, and then we were presenting those in front of the client.
David Moghavem (11:43)
What's Figma? Sorry.
Adam Rian (11:45)
Figma is a design system, is a system
David Moghavem (11:47)
Okay.
Adam Rian (11:48)
that you can design a software or design a solution, but without writing any code.
David Moghavem (11:54)
Mm-hmm.
Adam Rian (11:55)
so that that so that was the thing that we were using before before vibe coding. and then after
couple of days of working with this type of design software, then we were presenting something in front of the client and the client would say, okay, yes, I I like this or not. But now they can vibe code things themselves. And in a first session with a client, we can sit down and we realize, okay, this is what they really need because they vibe coded in this specific way. So vibe coding is very, very helpful to understand the requirements.
And to understand what exactly can what type of product can solve their problem. But some people think it's like 80 20 rule that with vibe coding you can get to 80% of the platform, but it's gonna be very hard to finish that 20%. But I think we have to think about it from different perspective. So vibe coding is great for prototyping.
And requirement gathering and make sure we understand the pain and how that pain can get resolved. So vibe coding is 100% good for this one. But for actual production, it's 0%. Because we cannot use any code that happened inside vibe coding to the production code. We have to regenerate all of those codes.
David Moghavem (13:24)
Why is that? Why is it?
Adam Rian (13:27)
Because it's gonna be very hard to maintain. These types of softwares,
David Moghavem (13:31)
Mm-hmm.
Adam Rian (13:32)
a lot of things are happening on the back end. If you want to push it into the production, we have to make sure we can maintain it. And there are gonna be lots of edge cases happening when someone uses the platform. And in the AI, those and specifically in real estate, those edge cases are most of the
The work is gonna be around those edge cases. For example, when
David Moghavem (13:57)
Mm-hmm.
Adam Rian (13:58)
you are ingesting lots of lees, there are lots of things that we have to pay attention to those leaves, like appendix, like amendment, like something change, or some someone cross a number and add another number by hand. So all of those things the the platform should be able to ingest and and clean the data and present it in the right way. And white coding is gonna ignore all of those things.
But at the same time, from an engineering perspective, when we want to maintain a code, we should be able to understand it. And AIs, they tend to do lots of coding, even on a text, if you ask AI to write an essay, so it's gonna do lots of tokens, lots of words, lots of like 200 pages.
And we don't even have time to read all of those things. And sometimes you don't need those 200 pages, you just need maybe two pages to say what you want to say. The same thing will happen in vibe coding. When you vibe code something, and when we look at the source code, lots of those codes are unnecessary. And that creates future bugs. So we have to remove them. And in order to remove them, we have a process. So there is a harness engineering style of.
Coding that we are using internally, and lots of startups are using, and through that harness engineering, we are able to remove hallucination and make sure all of those AI-generated code can be used in production. So that's the process. We sit down with a client, we see what they did in a vibe coding session on their own, then we use that for the requirement gathering. Then it's gonna be a process to learn from
From product perspective, we're gonna learn from what has been five vibe coded and then we're gonna build it from a scratch in a production ready environment.
David Moghavem (15:54)
So I want to go back to the concept because I think it's important of how vibe coding can only get you so far with prototype, but the fact that you have to continually maintain it in order for it to be productive ends up being a detriment to your own productivity. And so that's
Adam Rian (16:14)
Exactly.
David Moghavem (16:14)
where actus can come in. That's where enterprise software can come in and actually make something that you vibe coded into a real sustainable platform.
That you don't need to be a real estate guy that's part-time engineer anymore. Like I've fallen into that trap before where I've maybe vibe coded a solution and then I spend more time trying to maintain it to make sure it works instead of that doing the actual production of what I'm trying to achieve. And so for those of you out there in our audience who feels like they're good with AI or strong with some of these out-of-the-box solutions.
I'm sure you've fallen into traps where you're like, why am I spending half my day trying to correct this to maybe be more efficient later, only to have the software, the code break on an edge case? so I think we're entering that phase where we're feeling the limits a bit on these out-of-the-box AI solutions. And if we want to actually take AI to the next step in our own businesses, we need to maybe.
let go of the the keys or let go of our own egos and see what's out there and almost level up with the code so that it's sustainable and can lead to real results.
Adam Rian (17:35)
That's correct, David. And I'm really happy that you you mentioned that. And we've seen lots of these products that they failed because of exact problem that you mentioned. it's give you lots of good feeling that I was able to achieve all of these things in just a weekend. but the real test is the productivity gain that you mentioned.
If
we're gonna if we want to get to that productivity gain, we have to design a system that can replicate your existing workflow. And we should be able to measure the success of that specific workflow. If we cannot measure it, we cannot improve it. And that's the crucial part. Measuring the success of this type of productivity tool is important, and that should happen from the design, not after the fact.
David Moghavem (18:31)
So going into measuring success success, and we actually have been going back and forth with your team about what defines true success with the platform. Because for instance, like, you know, we're doing a lead gen platform, we'll go into it a little more, but booking a meeting may be different than booking the right meeting, right? And so how do you guys approach defining customer success with your guys' product?
Adam Rian (19:01)
so that's a great question, and that's gonna be through onboarding process. when we are talking with real estate firms, so there are a couple of important things that they are paying super attention to. One is the when they do the deal screening, they want they don't want to waste their time on a deal that they should not spend any time on.
So they want to minimize the amount of time that they're spending on it, not a good deal. And that can be solved with the right AI tool. Then we need to measure, like some people, they are so good and understanding is it a good deal or not, or should I go deeper? Should I ask for data room? Should I sign an NDA or not? Then without AI is just a couple of seconds for them. But some people they need to spend maybe half an hour.
And do a little bit more analysis on that. So these are different from one person to another person, from one firm to another firm. And we want to make sure this new AI workflow respects what they are actually doing right now. Then at the same time, that one person, which is very, very good, their judgment is amazing on understanding is the good deal or is a bad deal, how we can translate that judgment into real AI workflow.
When that specific person they have an intern, what are the training mechanisms between the senior person and a junior person? That training mechanism we have to apply it into this type of AI workflow. Then we have to actually measure the time, how long it's gonna take to successfully screen a deal or successfully underwrite a deal, and how long it's gonna take to
Make sure we are writing down a right IC memo. so these times are very important. But the real metric is if that deal was really good or there is something was hidden that we didn't know about it. And sometimes AI makes mistakes, the same way that humans does make mistakes. And since we are early on and building this type of AIs, sometimes
We have to wait and let's say a year from now or two years from now, go back and evaluate all of those decisions that AI made. And that's easier when you have digital teammates or AI workers compared to the human. Because then when we have analysts, there are lots of things that might change. People leave, people you you hire a new people, lots of institutional knowledge will leave the organization when people are leaving.
But when you build this type of AI system, everything has been logged, and every you can trace back every single decision to a specific judgment that that AI has been trained based on. And because all of things have been logged, then we can go back two years ago and ask a question why we made this decision and what we learned from it. So all of those things we need to design it from scratch before start building this type of AI agent.
So time to make a decision is one of them, but at the end, the quality of those decisions are important.
David Moghavem (22:23)
Right. And the only way to know the quality of those decisions is the iterate reiterative process of going back to previous judgments, seeing what the outcomes are, and continuing that iteration.
Adam Rian (22:36)
Exactly. And how we can build a system that AI can learn from those iterations. So th so these are the things that we have to build when we are thinking about a digital teammate.
David Moghavem (22:48)
Right. It's it's a new way of benchmarking, right? As you said before AI, when a company would reflect ways for themselves to improve their own customer success from a human capability, there were so many variables changing. It could be a employer or employee that's coming in and out, or it could have been a bad employee or an outstanding employee that would sway whatever the metrics were. But now that everything is digitized.
and logged you can use that information to continue refining and re reinventing the way that maybe a workflow can be done.
Adam Rian (23:27)
Exactly. And one example is like financial models. So you mentioned Bruce Kirsch from our team. He's one of the top persons building this type of financial model and teaching the financial model. And we are so fortunate to have him in our team. So we spend lots of time together in order to see how we can design an AI system that can evaluate a financial models. What's a good financial model versus what's not optimum financial model?
How we can improve that. But what is a financial model at score? It's an abstraction layer that's gonna abstract all of these complexities that real estate has and then abstract it to verifiable numbers that we can use those in order to make that decision.
So someone has in basic couple of I think maybe seven years ago or even more, they invented a discounted cash flow model. So that's something that the mathematicians think through and invite in order to help people to make those decisions faster. So these are the abstraction layer that's gonna help us to make those decisions faster. So number one thing for the AI workflow.
is make sure it understands all of those existing abstraction layers and work with them. When you have a financial model, which is pretty complex financial model, you have to make sure AI understands it and AI can work with it. Then at some point you're expecting AI to give you hints and give you recommendation to improve your financial model.
And that's that's the point that we are designing the system with the client for the AI to get enough feedback to provide a guidance to change the financial model. We are still not at the stage that led the AI to make those edits by itself. We are at the stage that AI will give the recommendation and a senior analyst make those edits.
David Moghavem (25:30)
When do you think we'll ever get to that next stage? If if ever.
Adam Rian (25:35)
I think we will get there in less than five years. And the reason I'm using that five years is we need enough data. We need enough, and and this this is not universal. Some firms they might get there faster, some firms they might get there later. What the things that's gonna determine how fast we can get there is the high-quality data. So we need to give AI the chance.
To make a decision and learn from it. But humans should override those decisions right now. But we can do retrospectively and say, if we made that decision, what would have happened? And all of those things need to get logged because retrospective analysis is very important when you are using this type of AI. And then this type of retrospective analysis will provide more data for AI to learn. So when
real estate firms, they start using this type of AI agent now, I believe less than five years it will get to the point that it can make significantly good decisions. Better than
David Moghavem (26:45)
'Cause there's enough runway
to make retrospective analysis
Adam Rian (26:49)
exactly.
David Moghavem (26:49)
by having that window of of a five year period. Right.
Adam Rian (26:52)
Exactly, yes.
David Moghavem (26:53)
Yeah. It's it's gonna be an interesting time where we almost give the keys to AI to make those decisions and almost trust AI more than we trust our own gut or feeling. And it's some will say we'll never get there, some will say we'll get there soon. And but I do agree that that window
is a good benchmark for saying we'll have the data now where the retrospective analysis will become so strong and the iterative process will have occurred so many times. Just
Adam Rian (27:23)
Yes.
David Moghavem (27:24)
like you have a entry level analyst who's been working at the company for a long time, you don't need to correct that analyst as much or ever when they do that
Adam Rian (27:33)
Yes.
David Moghavem (27:34)
task for the hundredth time versus the first time.
Adam Rian (27:37)
Yes. I want to clarify something. Trusting AI more than trusting ourselves is a very broad term. but if we yes, if
David Moghavem (27:46)
Right. It's loaded. It's a loaded term. Yes.
Adam Rian (27:49)
but if we change the scope of those type of decisions and and then we can make decisions and this type of prediction better and faster, like in a financial modeling.
for real estate screening. So if we narrow down this scope, then it's gonna be easier for us to look at the data and see what that decision means and how we can teach AI to make those types of decisions. I believe human will make decisions much, much faster with these AI tools even now. And as we are building this type of system, then human
We'll make other types of decisions that we still don't know what those decisions are. Because our job descriptions are changing. Even inside
David Moghavem (28:39)
Mm-hmm.
Adam Rian (28:40)
our firm, when we are hiring a new person, their job description is different from 12 months ago. So and those
David Moghavem (28:46)
That's interesting. That's interesting.
Adam Rian (28:48)
those this there will be some decisions that human is gonna make in future that's gonna be significantly better than AI, but we still don't know what those decisions are.
David Moghavem (29:00)
Yeah, it's almost like we are there's going to be decisions in the future that we're making that we don't can't even fathom making them today. Mm-hmm. Could
Adam Rian (29:08)
Today, ex exactly. And those
require those abstraction lay.
David Moghavem (29:13)
Right. Could be algorithmic decisions and just like seventy years ago before Excel was born, decisions were made completely different than seventy years later, where we're deep into Excel models and we're making decisions on assumptions and growth projections that maybe in a few more years from now we're gonna be making decisions that are don't exist today.
Adam Rian (29:38)
Exactly, exactly. Like the like the question is, can AI invent something better than Excel? Because
David Moghavem (29:46)
Mm-hmm.
Adam Rian (29:47)
we as a human, we invented Excel, which is an amazing tool. Now we have an AI which we are training right now inside Actus that can operate inside Excel, which is amazing. So AI that can operate inside Excel and do all of those calculations and help you make a decision, this is an amazing achievement.
But the next level of achievement is AI invents something better than Excel, which hasn't happened yet, but it might happen in the future.
David Moghavem (30:18)
Right. I want to get back to the practical use cases that AKTUS has been able to accomplish with their clients. Maybe give an example of one of the projects you guys have worked on that you in A enjoyed doing, but B saw a tremendous amount of success that you've achieved for the client by doing so that they wouldn't have otherwise achieved if they didn't use you guys.
Adam Rian (30:45)
yes, I can talk about a couple of examples. most of them are around document operation in
David Moghavem (30:52)
Mm-hmm.
Adam Rian (30:53)
in in real estate and investment management. So in real estate.
When you want to make a decision, there are thousands of the documents that needs to needs to be read, understood, abstracted, extract some data from, and then make a decision. And those documents are large and huge. And not all of them are financial documents. Some of them are legal documents, like zoning, the zoning legal documents, lease abstract, lease documents, and also the financial documents.
documents. And right now, when you drag
and drop all of those documents in out of the box AI model, they will get hallucinated. And we can talk about a technical limitation which is around the context window that we can talk about it later.
David Moghavem (31:50)
We
we could we could bring it up maybe like a just for the audience who doesn't know like simple, short answer, like why do they hallucinate, you know?
Adam Rian (31:59)
So all of these AI, the same as our brain, they have a memory limit. if you give them more data that you can fit into their memory, they can start hallucinating. they call that a context window. How much what's the amount of data that you can give to AI and what's the amount of data that AI can give back to you? So that's that's the context window. And
David Moghavem (32:26)
It's almost
like if you're using one chat on Claude to do all your tasks, then it's just gonna hallucinate or get it wrong and that's yeah.
Adam Rian (32:35)
Exactly, exactly.
When when you create a new session, you reset the the context. So the more context means you have to pay more because you have to pay more token to to to AI providers. So both from performance perspective and cost perspective, we have to do context management. there is another part of this type of AI that although they are claiming
their context window can be up to one million token, which kind of equivalents to the keywords, but only 10% of that context window is effective.
David Moghavem (33:20)
That reminds
you that reminds me of the ice stat that your brain you only use ten or twenty percent of the capability or capacity of your brain. Yeah.
Adam Rian (33:28)
That's similar. Yes, that's
that that's something similar. So when you are reading a novel and then a week later you don't remember everything. And most of the time you forget things that happen at the middle. Most of the time you remember a beginning of the novel and end of the novel. And interestingly, AI has this the same type of behavior. When you give it lots of data, AI tends to forget things that it's at the middle of it.
David Moghavem (33:55)
The similarities between like the human brain and the way AI h remembers things. It it's incredible. It's it's yeah.
Adam Rian (34:03)
It's fascinating. Exactly.
It's fascinating. so now the same way that when we hire someone or when we are educating someone, we try to give them the right amount of data at the right time, not everything at once, the same techniques that we have to use for AI.
And that's what Act as AI is doing with a client. So when we are onboarding a client, we make sure we build and customize a context engineering platform for their need to make sure AI sees the sees data that they they need to see and when they need to see.
And it's sometimes it's very important to understand there's some part of the data that is better for you to hide from AI, because otherwise it will hallucinate. When you're talking about the finance, what part of the legal document that you have to show to the AI? So these types of things are are make the AI system more robust. And one of the the sucks success projects that we had in the past.
Was related to real estate investment decision, which we ingest legal document like zoning, lease document, and a financial document, and thousands and thousands of those documents. So on a back end, we were able to build an engineering system that slice and dice data and make a decision. The AI, which in this example is large language models, an LLM, like Cloud or OpenAI, they see what data and when to see it.
Then
they provide an answer. One of those input data to the system was a real estate financial model with 60 plus tabs, pretty complex comprehensive underwriting model. That was another input to the system. So right now, if you if you upload a comprehensive financial model into the cloud or OpenAI, they they hallucinate a lot.
Complex to understand that financial model. We have to build an engineering system to break those problems in the smaller piece and make sure AI understands it, then test it to make sure AI passed those tests, then use it in the underwriting process. So that was a summary of one of the success projects that we had in the past.
David Moghavem (36:35)
Right. I think it's important understanding you you went deep into a talk context management, where you can't just that's a slang word I've heard is one shot. You can't just one shot for one of these out-of-the-box AIs to do everything for you. It's gonna hallucinate, it's gonna make errors. It feels like it has to answer. It's almost like and it answers with confidence, right?
Adam Rian (37:01)
Yes.
David Moghavem (37:02)
And so you don't.
It doesn't say I'm not sure, but I think it's this answer. It answers in confidence because that's just the way it's been trained to, and that leads to a lot of errors. And that's where you guys can come in and point out the holes in AI and the fallacies and be able to optimize for it to be right and certain.
Adam Rian (37:25)
Exactly. So AI is an amazing tool. We need to use it in a right way. And that right way is this context engineering platform. And we call it when you harness OpenAI, when you harness cloud, then it's it's it's it's gonna solve those those those problems for you. There is no magic, there is no
one-shot learning. It's always requires back and forth. Human is not gonna happen for human it doesn't happen one shot. For AI is not gonna happen one shot. And these type of decisions in f in in investment is pretty complex. We have to take our time and make sure AI get it and and and we test it before we give it the specific tasks.
David Moghavem (38:07)
Right. So one thing we're using with you guys is we're building a lead generation platform for third sourcing third party management. And sourcing can be for third party management, can lead to acquisitions, loan sourcing. But the idea is that the way in which we're going about it is not your simple email saying, Hey, are you looking for management? Here's try on living. It's something that's a little deeper. It's actually ranking the properties.
whether it's close to our own portfolio that is performing well or other properties in different markets. And we're using as an anchor point to source and power rank the properties from you know best to worst, having different approaches for all of them and being able to have AKTUS actually thoughtfully reach out to prospective clients and see where we can actually add value, not just do a spray and pray.
And you can only imagine how trying to vibe code that is a disaster. You can maybe vibe code a component of it, maybe a type of deliverable, or maybe you can pull a certain contact, but to do it at scale is just a whole different project. And I think that's where for us we saw so much value with you guys that trying to vibe code that would be a disaster and having you guys on board has been has been great. So I'm excited to see where.
Where that goes.
Adam Rian (39:37)
Absolutely.
Thank you, David. And that's a great example of how digital teammates work. Let's say if you want to hire an analyst to do that task, then all of those things that you mentioned right now, you will let them know during the onboarding or even during the interview that these are the things that should happen. We should not spray and pray. We have to customize it. These are the ways that we have to do the search. So all of those things that you just mentioned.
These are the learnings and guardrails for AI. Now we're gonna let AI to make some decision, to do the sorting, to draft an email. So, as you know, we are not gonna let AI to send email yet. We're gonna let AI to draft the email. And those the draft, the input to those drafts is all of those judgments and knowledge.
And all of those documents that need to be ingested, and the AI should read them, understand them, the financial, the location, the zoning, and make sure that's the right prospect. And then, based on all of those, write down a customized email. Like, hey, I know you are doing specific things. We are your neighbor in that specific area. We did all of those things. And all of these emails are gonna be customized and changed.
Based on the prospect. And we are not at the point that's gonna let AI to send all of those things. The AI is gonna just draft them and make sure someone like you is gonna read them and review them. And then based on those feedback, you might say, okay, 90% of them are amazing. I'm gonna hit send, but the 10%, the AI should go back and redo it.
David Moghavem (41:20)
Right. And I think it saves the time. So now when you're sending that email, let's say the person doesn't respond, deletes the email, which is happening all the time.
Adam Rian (41:29)
Yes.
David Moghavem (41:30)
Rather than just leaving it at that, now you have a little bit of context of yourself of who that owner is, where they've transacted. You can pick up the phone, you could call, you can leverage some of your own connections. So it saves so much time from doing the lead cracking and the email writing.
and the and the sourcing to now you have that time and you can dedicate it towards finding the right relationship, talking to the right people, leveraging your own connections. And I think that's the next step for humans is as you said, our decision making and our lives are going to look a lot different in the next couple of years to where it was just a few years before. We're going to be doing less of this type of robot AI work and we're going to be doing more
Adam Rian (42:16)
Yes.
David Moghavem (42:17)
of this
Human leveraging work.
Adam Rian (42:19)
Yes.
Yes, yeah. And some people might think, okay, we are building a new CRM. No, that's not the case. It's it's a digital teammate. It's like
David Moghavem (42:29)
Right.
Adam Rian (42:29)
someone who are going to execute those type of tasks under your guidance. Sometimes those tasks are evaluate this OM, run this financial model on that data room, and give me the report. Sometimes the task is find these 20 prospects in this specific area.
So these are the tasks that our AI system is gonna help you to execute, not a new tool, not a new software, not a new application. And that's the difference that I want to make sure our audience understand this is a new paradigm of executing tasks. And in near future, I believe these type of AI agents, they demand their own application.
So all of these applications that we had in the past, they've been designed for humans to operate. Now we are building digital teammates together. And those digital teammates, they demand a specific type of tools. And those tools, we don't know what those tools are yet, but we have to work together in order to understand those. And then maybe sometimes we have to go and purchase a specific tool for our AI agents.
David Moghavem (43:45)
Yeah, I mean, like one thing that for me that rings the bell is like API keys, right. Where a human wouldn't need an API for their own. They could just go on to the UI and look at the data. But I think for efficiency's sake, having the right API keys and integration, I think that's the next step. I was actually saying how some of the tools we use that don't use or share their API key, I think they're behind. They're gonna end up getting lost in the sauce with what's going on.
where if you're leveraging AI, a tool that doesn't share their their API or share their tools almost can be become obsolete.
Adam Rian (44:25)
Exactly, exactly. So in future, we are going to have all of these digital teammates communicating with each other and working with each other. The the people that embrace this type of new transaction will win the market. And we have to design the system, we have to design a software that embrace a communication between different AI agents.
David Moghavem (44:49)
I wanna wrap it up with a question that I wanted to ask you. What's one thing about AI and real estate that's overhyped and one thing that's maybe underhyped?
Adam Rian (45:01)
So I would say most of the thing, most of the new technologies are overhyped in short term and underhyped in long term. So if we use internet for short term, for that specific couple of years, it was overhyped, but long term it was underhyped. So I believe specifically for for real estate, just drag and drop and one shot dealer screening is overhyped.
This is not something that people can rely on in a long term. Lots of people are talking about it, but I don't think it's gonna work in in in in in real world. But in long term, even that's gonna be under hype. and by long term, I think we have to wait maybe 12 months, 12 to 24 months for a couple of those.
overhyped, like out of the box one shot that you mentioned just one shot dealer screening. So these type of things will go away and customized solution for dealer screening, AI powered customized solution for dealer screening and asset management will replace those type of one shots. But overall I believe AI, not just for real estate, as a whole, as a new technology, is overhyped in short term, is underhyped in long
David Moghavem (46:24)
Well said. I think I agree with that. and so it'll be interesting to see where things evolve. It'll be very interesting to see how AKTUS evolves over over the next, you know, years and decades and and this in this cycle. Where what do you see next, I guess, for actus in these next twelve to twenty-four months as we go from a overhype to underhype type of environment with AI.
Adam Rian (46:51)
We
believe our customer will design their new org charts based on Act as digital teammates. So one of the things that AI will change the industry is the org chart itself. So why we have these specific org charts right now, because of our education system. We went to a school, we got architectural degree or finance degree or legal degree, and based on that.
The whole enterprise has been designed around those specific skill sets that we learned through school. AI is changing the whole education industry. And multidisciplinary's task is gonna be more achievable.
In near future. And those multidisciplinaries is gonna change the education system, is gonna change the skill set of the people. And in near future, we have to rethink how we are designing our orch chart. Do we need a marketing department separate from finance department? Or those orgs will be around actions, will be around the task or job to be done, as opposed to a skill set of humans. Our vision.
Our roadmap for Actas is in near future, is gonna influence the org chart of of enterprise and and they are gonna use Act as digital teammate to do the operation.
David Moghavem (48:14)
That's that's a a world where it's gonna be completely different, right? Digital teammates, having you know, basically AIs. I mean, I don't wanna say replace because as you said, I do think the human is gonna just also evolve to what they're gonna be doing, but it is replacing the jobs today. But I think it's also creating the jobs for tomorrow. So it's
Adam Rian (48:39)
Exactly.
David Moghavem (48:40)
gonna be interesting to see how AKTUS plays a role in that. Adam.
Really great having you on. Loved how we got theoretical with AI and practical with AI in real estate. for those of you who couldn't tell, Adam is also Persian brother. So Adam, Khali Khoshumadid,
Adam Rian (48:59)
Mm-hmm.
David Moghavem (49:00)
great having you on and looking forward to linking up soon in person and continuing working with the actus team on our project.
Adam Rian (49:08)
Absolutely. Thank you so much, David, for having me.
David Moghavem (49:10)
Awesome. Thanks.
Adam Rian (49:12)
Thank you.
