How I work with AI.
I use AI in two places: to run a program, and to guide a build. In both, the agents don’t decide. I decide what to make and what to cut, write it down, set the tests that prove it, and check the numbers that come back.
On a program
Requirements
Every contract requirement traced to its epics, designs, rules and tests. A link the tool inferred stays marked as inferred until a person confirms it.
Planning
One board for five increments. Every metric carries its source and the query that produces it. A gate with no data shows a dash, never green.
Reporting
Reusable AI skills rebuild the delivery report and roadshow numbers from exports, and reconcile increment figures against program-wide ones.
The team
I am the AI resource for the business analysts. On the shared codebase, agents follow the team’s loop: intent, plan, failing tests, then a verifier and a human review.
On a build
Decide
I own the problem, the user and the cuts. Every open question is a decision with my name on it, and exclusions are written down with a reason.
Write it down
Talk, document, build. Plans, specs and numbered requirements come before code. When 34 plans disagreed, I merged them into one.
Prove it
Tests are the referee, not the model’s word. A manual became 180 tests. Two engines had to agree on 13,225 values across three compilers.
Review and reverse
Design critiques, mock-ups before builds, and decisions reversed in the open when their reason disappears, with the reason recorded.
Calls I made
- The PI dashboardA gate with no data shows a dash, never green.
- The PI dashboardMachine-suggested epics stay marked as proposals: they matched the team’s own only 20% of the time.
- The RTM toolAn upload shows what will change before it writes anything.
- The PI dashboardInterviewed every workstream lead before planning, so ten vendors plan from one script.
- Job SynthesisA ledger of confirmed skills, so AI tailoring cannot invent any.
- idMShipped as an instrument with MIDI out, because Ableton rejects a MIDI-effect VST3.
- idMMoved the engine from Rust to C++ a day and a half in, when the reason for Rust was gone.
- Audio Edit & TagBuilt a sixth stretch engine, tried it, and cut it the same day.
- idMKept MIDI file import out until a real workflow needs it.
- The AI sound labLocal agents work against hidden tests; the bridge re-runs the tests itself.
My AI years
- 2022
Ran Stable Diffusion on my own machines through Automatic1111, and used ChatGPT in its first week.
- 2023
Made AI video with Deforum, steering the motion in real time.
- 2024
Led delivery at a stealth AI startup fine-tuning Gemini, ChatGPT and Claude models on customer data.
- 2025
Started building apps with AI: visualizers, starfields, games.
- 2026
Built four music products on shared engines with AI agents, a home AI sound lab, and the AI tooling for a ten-team program.