Build first.Talk after.Repeat.

I’m Joel Jeon. I build infrastructure for capable, verifiable AI agents: specialists, capabilities, and systems that check real outcomes.

joeljeon7@gmail.comGitHub ↗
THE EVIDENCEBuilt, then
verified.
CURRENTDynamic Agent SpecialisationPrivate local prototypeTESTEDThree fictional rolesSeparate outcome verificationDISCIPLINE250+ local checksNo unproved upgrades
CURRENT WORK

Dynamic Agent
Specialisation.

DAS takes a bounded business role, builds several complete specialist-agent configurations, tests them against independently checked outcomes, and recommends the strongest safe option only when the evidence justifies switching.

The private local prototype has been exercised across procurement, SaaS support, and CRM/revenue operations, with more than 250 automated checks. No real customer system has been activated. I also work on Growth/GTM and growth-engineering experiments at Tsenta (YC S26).

Read the technical note ↗
6
CAPABILITY FACTORY / SIX DAYS

A missing ability should not end the goal.

Capability Factory lets an agent diagnose a missing ability, acquire or reuse the smallest permissioned capability, independently verify the result, and resume its goal. It is a working local pilot MVP for constrained HTTP APIs, tested only with fictional data.

Read the technical note ↗Visit Capability Factory ↗
FUTURE ARCHITECTURE

DAS and Capability Factory are separate products today. Longer term, a bounded Fleet Brain could coordinate specialists against company goals, call for new capabilities, verify external outcomes, and stop for human authority.

Build specialists.
Give them abilities.
Coordinate the fleet.

The Agent Fleet Brain is a plan, not a built product. YC at 16 remains my personal target.
PAST WORK

Things I built.

weld. first, then three hackathons under pressure.

01First product / Roblox talent marketplace

weld.

A marketplace for Roblox developers and studios. I taught myself to code to build it, shipped the landing page on day one, then killed it when the business model did not hold.

0236 hours / 12% acceptance / £10k top prize

Pop the Bubble

A few weeks into coding, I built Synapse, an AI growth-analytics tool, for the largest hackathon prize in London. Youngest person there. I barely made it through, but I finished.

03Lightfern track / ~150 people / second place

GTM Hack

Built the Voiceprint Funnel with my friend Jawad. We were the only two people in the room who were not university students.

04Solo / two sponsor categories / second place

LicenseTrace

An agent that traces hidden open-source license contamination through a dependency tree and proves the exact path. £300 cash and £15k in ontology-AI credits.

FOUNDER STORY

It started with
a gap in Roblox.

Thousands of developers and studios were building on it, with no real way to find each other. So I decided to build weld. ↗

I had never written a line of code. I taught myself because the thing needed to exist, then shipped its landing page the same day I started coding.

The first code I ever wrote.

weld. never found a business model, so I killed it. What survived was more valuable: I can build the thing I imagine.