How I Built an Ultra-Low Latency, Local-First Interview Copilot with Electron, React 19, and Groq Whisper Turbo

I have to come clean. Speaking at APIWorld this past week wasn’t actually my first talk acceptance....

One month after the queue essay: findings file themselves, rounds are judged on the contract, the morning brief reads like a PA's note, and a refused write is never silent.

The permission prompt only fired when I was watching. A hook now refuses force pushes, shared-tree resets and root deletes in every mode, and redirects the session instead of waiting for me.

296 puzzle cards, each judged by a model that sees the story but never the brief, the finding or the old card. Eleven waves, and the brief turned out to be the lever.

The voice engine flags it, the site refuses it, the drafter is forbidden it, and 117 instances were rewritten by hand. Then the gate blocked my own deploys for a day.

A backend team shipped a local-first coding agent that kept repository secrets on each developer...

The lab was quiet except for keyboards. A teaching assistant walked the aisle of a Saturday workshop...

Open any 'agentic AI' repo trending on GitHub this month and you'll find the same shape: a while...

Same base model, same sampling, entirely different coworker. The diff was 40 lines of English, and here they are.

On August 1 I published a report whose headline finding was that frontier LLM API prices are...