DWG EM-2026 · Sheet 01 · General assembly
I build the jig.
AI builds the part.
More than a decade of infrastructure and system architecture, built by hand. Now I brief AI like a teammate, box it in, check its work, and sign for what ships.
The tighter the jig, the freer the hand.
The slower the proof, the faster the ship.
NowBuilding Pradini. Chauffeur-driven car rental for Lagos & Abuja.
Sheet 02 · Hands · 2013–19
Self-taught, in Lagos. Freelance web work from 2013, then Andela’s bootcamp in 2016: a new language most weeks, learned by building with it.
When I don’t understand something, I build it. My first repo was a command-line banking system. After that: a JWT library, logic gates built up from NAND, a 2D physics engine.
If you want to understand it, build it.
Sheet 03 · First crews · 2016–22
At Andela I joined client teams in stacks I hadn’t used before, and had to be useful fast. As a bootcamp facilitator I took more than 30 candidates through the last stage of Andela’s application process; over 90% were accepted. Then I led 6 developers on 4 client apps, with reviews and mentoring for each of them.
On a remote contract I helped split a large app into smaller services and looked after 9 of them in production. As a first engineer, I moved a team off deploying by hand. Then I co-founded PingPong as CTO: I built the first version alone, then led a team of 5.
Code takes the shape of the team that writes it.
Sheet 04 · The Recorder · 2020–26
Teams across time zones needed to talk without being online at the same time. I kept following that problem through five versions: PingPong’s async video, then my own recorder, native capture in Swift, Captur in Rust across 9 meeting platforms, and now Nzuko, which is self-hosted and transcribes on device.
Each step went a layer lower, but only when the need was worth it: better audio, less memory, and finally privacy, so nothing leaves your computer. What changed at the end was the cost. Agents made low-level code cheap to build, so the need and the timing finally lined up.
Go lower when the need is worth the cost.
Sheet 05 · Repairs & reach · 2022–26
A rewrite had been stuck for months because it was planned as one big switch. I turned it into a gradual move, one part at a time, and it reached production in about a month. Repair beats replace.
Then 3 years on job boards for investor networks: boards they could edit themselves, the admin portal and design system behind them, and AI matching with search in plain words.
After the acquisition I built two things. Warm-introduction paths: the best way to reach someone is through a person you already know, so the search finds those friend-of-a-friend routes, ranks the strongest and answers almost instantly. I kept it in Postgres, with the reason written down. And hiring with AI agents: contact lists enriched instead of duplicated, and agents that work out the ideal candidate, find and contact people, set up applications and grow lists.
Design for the load you have, and write down why.
Sheet 06 · The Crew · 2026
A small team can’t win by typing faster. Now I lead one engineer and a crew of agents, and my job is making that team work like a much bigger one: a clear brief for every change, 3 lanes in isolated worktrees, one fresh agent per ticket, merged by script.
About 20 skills and 27 CI workflows check the work before it merges.
If execution is instant, strategy is everything.
Sheet 07 · The Guard · 2026
Dozens of money bugs passed review while looking correct. Review alone couldn’t catch them, so I built the Guard: a check with no model in it.
Every guard has to be seen failing once before it’s trusted.
A guard that cannot fail is worse than no guard.
Sheet 08 · Where I’m heading
Let the model write, make the guard check, keep the person deciding. That’s what I’m building next, as Datum.
Read the heading sheetSheet 09 · More sheets