What I built. How I tested it. What I had to fix.

Lumina is a character people talk to in a voice call. She hears you, decides whether you were talking to her, answers in her own voice with a moving avatar, remembers you, and plays games on stream. Her personality is a language model I fine-tuned and run on my own GPU; small, fast classifiers around it make the yes-or-no decisions. What matters to the person talking to her is how quickly she answers.

• 1.74 sfrom the end of your sentence to her first audio (median of 20, was 2.09 s)
• 0.27 sto decide whether you were talking to her; the step it replaced took 0.74 s
• 46 / 50memory-router calls correct through the live path, 0 false recalls
• VoDshe completed Minecraft on hardcore with no human intervention; the recording is public

1. 00
   Listen

   Discord voice arrives encrypted per speaker. She waits half a second to be sure you have stopped, then turns your speech into text in about a fifth of a second.

2. 01
   Remember

   Her three nearest memories about the speaker come from a local vector store in about a tenth of a second. I removed a one-second cloud call that only decided whether to look.

3. 02
   Gate ‖ persona

   Two things start at once: a quarter-second check decides whether she was being spoken to, and her reply starts being written. She speaks only when both are done, so the slower one sets the pace.

4. 03
   Speak

   Her voice starts a quarter of a second later, streaming the first sentence while the rest is still being written.

5. 04
   Lips & face

   The same audio plays on my machine, and that local playback drives her mouth, so lips match what you hear. A last quick check picks her expression.

One agent writes down a status or raises a blocker. The next agent, cold, on another machine in another repo, reads it and picks up the work. Nobody relays the message. It is a Next.js app and an MCP server with a knowledge graph, a mirror of the code, and a review queue a member agent cannot approve its way around.

• 48tools an agent can call, across notes, code, work and review
• 155tokens for a whole project briefing, when I measured it
• 2 of 4rows my own benchmark loses to plain grep, re-run after I fixed a leak in the test. Both are in the README.
• RETRACTEDmy "33x cheaper" claim. It compared two different questions, and I corrected the README the same night.

1. 00
   Get briefed

   An agent that has never seen the project asks the vault where things stand. It gets a short summary built from real notes, about 155 tokens when measured, with any proposals waiting for review.

2. 01
   Read the mirror

   The vault keeps a copy of the code, synced from git. The agent reads the real file from that copy. It needs no clone and no access to the repository.

3. 02
   Suggest a change

   The agent cannot push, so it files the change as exact before and after text, with a written reason. The reason is saved as a note that outlives the review.

4. 03
   Review & apply

   Only an account with the owner or executive role can approve, and approving changes no code. The owner's agent takes the approved result, writes it into the owner's own checkout, and pushes.

5. 04
   CI mirrors

   The push starts a workflow that copies the new commit into the vault. The suggestion then shows as applied. The vault never writes to git.