PORTFOLIO — 2026LANDING · DATASHEET FORMAT
QI YANG
FULL-STACK, AGENT-FIRST
I build software with a four-agent fleet — planner, coder, reviewer, judge — under CI gates. I'm the human gate. What I ship is trace-backed; what I can't trace yet, I don't show yet.
Position
THE OPERATING BELIEFUnverified output is worth zero.
Ten years across finance and engineering taught me the same lesson twice. A backtest nobody can reproduce and an agent patch nobody reviewed have the same market value: zero. The failure mode isn't bad output — it's output that looks finished.
So I don't work harder at judging by eye. I work where verification is structural: CI gates machines can't charm, reviewers that must cite a line to say no, and one human who signs before anything moves.
System
HOW I KEEP OUTPUT HONEST| R·01 | Reviewer is never the author | The agent that wrote the code can't be the one who approves it. Separation of duties is enforced by the task schema, not by etiquette. |
| R·02 | Every rejection cites a line | "Check quality" produces approval theater. The review brief lists what must be refused, and a no without a line number and a rule doesn't count. |
| R·03 | Judges vote blind | Three judges see the artifact, not the authorship. Majority rules; a 2–1 split is recorded, not averaged away. |
| R·04 | One human signs, every time | No merge without my sign-off — the gate is the slowest step on purpose, and the signature is logged with the trace. |
Why rejection is the feature: a reviewer told to approve will approve. The loop that looks like friction — rejections, re-runs, re-votes — is the only part of the system that produces information.
Live status
THE SITE IS THE SAMPLEContact
SEATTLE, WA · OPEN TO REMOTEAvailable for: senior full-stack and AI-orchestration roles, and the occasional unfixable problem. Evidence ships with the works — the project index opens when the traces are real.
// FASTEST CHANNEL: GITHUB — USUALLY SAME DAY, PDT