I run AI agents like an engineering org.

What it delivers

Velia went from its first commit to public in 24 hours. When a sailor on Reddit said its waves sounded metallic, the sound was diagnosed, rebuilt, compared side by side and shipped the same day. At my day job I work AI-native at the scale of an engineering org.

In that month: a browser sailing game with real wind and water physics; a robot's CAD, firmware and physics simulation; an eye engine in Python with a C port held to parity; production web launches.

The org

• One integrator, many builders. Each product has its own agent, each task its own branch; an integrator merges, checks and ships.
• Rules are code, not advice. A commit is refused without its design doc, or with a private path or a key in it. No agent, and not me, skips a hook.
• Every mistake becomes a one-line rule that every new agent loads, so it is not made twice.
• I change the structure when it stops paying. I retired a six-role agent team in one day when coordinating it cost more than it delivered.

Proof, not claims

• A test counts only after it has failed on code broken on purpose.
• Results are read from the system, not from an agent's report. A fall is what the simulator measured; a look is the pixels of the frame.
• On Velia, 971 tests and 10 pixel-matched frames run after every merge.

Taste stays human

• Anything a person sees or hears merges only after I approve the before and after. Agents never judge beauty.
• I judge on the real thing, the phone in my hand or a full-size crop, never a summary.

Cut fast

• Measure, then decide. A faster start for Velia saved 1 to 2 seconds but added 3.3 MB to download; I did not ship it.
• After five rounds without progress on Snoot's fur, I stopped tuning, painted the target look and rebuilt to match it.