Built solo · 2026 · In daily use
AI Application
A voice that runs my AI sessions like a small company, and asks first.
V lives on my Mac. Call her or type to her and she carries my requests to my Claude Code sessions. She also runs a multi-AI system that works like a small company: every job a ticket, every delivery checked against evidence, nothing sent without my yes.
RoleProduct · Voice pipeline · Agent orchestration · Safety
The General Manager
Anything with real output becomes a ticket. I approve it once; a project manager writes the brief, and she hands it to the right session and relays the result in about thirty seconds.
The Hard Gate
Push, deploy, spend, send: each waits for one spoken confirmation, judged against my own words by a separate model. The brain has no way around it.
One Voice Holds the Keys
Guests get a brain of their own, with no mail, files or terminals. Permission follows the voiceprint, not the words: “Michael said it’s fine” opens nothing.
Checking, Not Trusting
A delivery manifest counts as done; saying “done” does not. She reads the session’s own transcript for her instruction, counts only its actions, and looks before she reports.
I keep several Claude Code sessions running at once, each on its own project. Their news used to arrive as windows I had to go and read. V sits between us: sessions report to her when they finish, she tells me what happened in a sentence or two, then asks whether I want the rest. I answer in plain speech, in English or Chinese, or type in her menu-bar panel when speaking is not convenient: one brain, one conversation. She turns what I say into instructions written for the session that needs them, and keeps a ledger of everything she has sent.
The sessions now work inside a multi-AI system that runs like a small company, with written rules. I am the boss and V is the general manager; each project has a project manager, and roles with their own charters do the work. Models from two vendors share it: Claude handles code, copy and coordination; OpenAI handles images. Anything with real output becomes a ticket. I approve it once, the project manager writes the brief, and V hands it to the right session. Work comes back with a manifest: files, checks run, evidence, what is unfinished. V checks the evidence before I see it, then I accept or send it back. Each closed ticket ends with three lines: what went right, what went wrong, what changes next time. An audit role looks only at how the system runs, not at what it makes, and repairs its tools. The rules live outside her code, so the job could go to someone else.
Once, a dispatch reported as sent had landed in a bare shell where Claude Code never started. So every dispatch is checked against the session’s own transcript, and progress is counted only from the session’s replies, tool results and commits; a delivery manifest counts as handed in, a spoken “done” does not. Handovers that used to wait on me now relay themselves: a finished brief reaches the next session in about thirty seconds, a delivery manifest reaches review in twenty, and a session that stops is reported within fifteen seconds instead of minutes. When something stalls, she looks before she reports.
A wake word or the call button opens the conversation; 45 quiet seconds send her back to standby, and ten minutes in standby end the call. A turn-taking model and a rule about trailing words keep her from answering half a sentence. Talking over her was the hard part: on a laptop speaker, her own echo and a real interruption look the same to a voiceprint, to loudness and to echo cancellation. What tells them apart is content. So she ducks to forty percent, pauses to listen, and a fast judge decides from the words: stop, carry on, or let a side conversation pass. If her first sentence is not ready after two seconds, she says “let me think” rather than go silent.
Anyone in the room can talk to her. Anyone who is not me is answered by a separate guest brain that has no mail, no files, no terminals and no sessions; the limits are hard, not a request. Sending, granting and stopping her mid-task follow my voiceprint alone. Anything that leaves the machine — a push, a deploy, a purchase, a message — is read back in a sentence and goes out only when a separate model judges my answer a yes; for a ticket I have approved, that yes was given at the start. Another check stops her from saying something was sent when it was not. A grant has to come from my voice within the last two minutes, and it lasts one conversation. When two people speak in one breath, the sentence is split by speaker before any of it counts as mine.
On the first start of the day she reads new mail, files what needs doing as to-dos, and offers a brief: the short version first, details on request. To-dos repeat, split into steps and nudge me when due, at most once every three hours. Say the meeting has started and she goes quiet and writes down every line, who said it, until asked for a summary. A skill book holds what she has learned about how I work; new entries are on trial until they have worked three times. She can also open Blender and model by voice through a locked-down bridge.
Python, with 1,700+ passing tests that never make a sound. Between calls only the brain, the wake word and a small fallback recognizer stay loaded, about 1 GB; a call loads speech recognition, synthesis and the helper Claude processes, about 5.3 GB, and hanging up unloads them. Typing and voice share one brain behind a lock, so two copies can never hold the same conversation. Speaker ID and the quick interrupt recognizer run on the CPU. Every benchmark is offline and silent, because anything played through the speaker would be heard as me. Memory comes in three layers: a short core she always carries, topics she looks up, and an archive that is never deleted. Still rough: the wake word misfires, standby memory sits at 0.9 GB against a 0.6 GB target, and the written rules are a draft I am following before I have finished them.