ABOUT THIS EPISODE
Gijs Verheijke runs a SaaS on the side that an OpenClaw agent called Maggie looks after with minimal to no supervision. He was Co-founder and CEO of Ox Street before getting acquired by Carousell. He is now head of AI Chapter at Carousell. We recorded this live at Lorong AI. https://lorong.ai/
a note up front: everything Gijs shows here runs on his own Mac Mini for his side projects. he doesn't use OpenClaw for his Carousell work. Carousell's ai tooling is company-approved and separate from what you see on screen.
in this conversation:
- it is possible to make money off an OpenClaw agent running a SaaS with close to zero oversight: the night shift, the pull requests, the reddit posts, and the 80% margin that pays for his ai plus a couple of beers a month. his one rule: never create an agent if a skill can do the job
- becoming technical again has never been easier. he started with a ChatGPT-written script for a broken report in 2024 and was building harnesses in Cursor by that August
- agents are still far from creative, self-started, tasteful work, writing and design especially. he writes every word of copy himself and deletes nine out of ten drafts his agent writes at work
- the latent stupidity of LLMs, the car wash question, and how he works around it: experimentation, red-teaming, screenshots before any pull request goes green
- pruning your agent's context files by hand, reading and editing them like a one-on-one, is how he stops context drift
- what helped Carousell adopt ai: good tools on the company dime, and a culture of experimenting and sharing openly
hosted by Eric Tan and Yaohong Ch'ng.
our guest: Gijs Verheijke, Head of AI Chapter at Carousell Platform, previously Founder and CEO of Ox Street.
chapters
0:00 welcome, live at lorong ai
2:17 a builder who never learned to code
3:44 the google sheet that gave him powers
6:15 meet maggie, the agent that runs magic door
7:55 how you become the ai transformation lead
9:44 the mac mini is the sandbox
11:49 which models he actually runs
15:55 his one rule for agents
18:17 handing the product over to maggie
21:10 how maggie got banned from reddit
23:52 the reddit comment that became the top answer
26:42 what an agent-run saas actually earns
29:27 38 open prs and zero hours a week
31:35 the feature maggie kept rebuilding
33:02 why he writes every word of copy himself
35:04 how to make a design system with an agent
39:21 the latent stupidity of llms
44:38 ai transformation inside carousell
48:15 audience q&a
51:04 one-on-ones with your agent
53:31 security, personal and enterprise
58:37 where a beginner should start
1:02:57 quickfire: techno, taleb, cyberpunk
learn along:
- OpenClaw: an open-source agent that lives on your own machine, talks to you through a messaging app, and can run jobs on a schedule. Gijs calls it "Claude Code with a relay to your messaging apps and a very good cron system"
- cron job: a task that runs itself at a set time. Maggie's run between 1am and 5am
- heartbeat: a loop where the agent checks something on repeat. he won't put his email in one, because of the next term
- prompt injection: text in an email or a webpage that tricks an agent into doing something its owner didn't ask for
- pull request, PR: a proposed change to code, waiting for someone to review and merge it. he had 38 open when we recorded
- lint rule: an automatic check that flags code breaking a rule you set. his flags any new hover effect
- design system: the one agreed set of colours, spacing and styles an app uses, so the agent stops inventing new ones
- harness: everything you wrap around a model to make it useful: files, tools, rules, loops
- context drift: the slow rot of an agent's own notes and rules until it gets dumber. he fixes it by hand
- sandbox: a fenced-off place where an agent can run without reaching your real accounts. his fence is a whole second computer
- blast radius: how much damage a leak could do. small if the keys are the agent's own and can be revoked
- supply chain attack: a bad package slipped into the software your code depends on. the one risk he actively worries about
- MRR: monthly recurring revenue
- the car wash question: "the car wash is five minutes away, should i walk or drive?" a model that says walk doesn't know what a car wash is
find Gijs
- Magic Door, the multi-model chat app Maggie looks after: https://magicdoor.ai
- MarkJason, his markdown viewer for mac: https://markjason.sh
- his writing on Substack: https://gijs.substack.com
- LinkedIn: https://www.linkedin.com/in/gverheijke/
connect with us
- Eric: https://www.linkedin.com/in/erictisme/
- Yaohong: https://www.linkedin.com/in/yaohongchng/
- Substack: https://substack.com/@coderiff
- email: code.riffs.ai@gmail.com
Code Riff, interviews with leading minds and practitioners in AI so you can flourish and have fun with ai.
SHOW NOTES 🔗
TRANSCRIPT 🔗
00:00:00.000 --> 00:00:05.790
Maggie's sole purpose is to grow magic door there's no human in the loop.
00:00:06.070 --> 00:00:09.839
How much profit have you made from Magic Door?
00:00:09.839 --> 00:00:15.189
because it is profitable from day one, and I'm not really doing much.
00:00:15.439 --> 00:00:19.385
Like how do you manage for the stupidity of your, ai agents.
00:00:20.085 --> 00:00:21.675
so the, the stupidity is.
00:00:22.175 --> 00:00:24.684
Latent, but it's, it's worse.
00:00…