Mini-me is an agentic twin trained on how you decide — your past calls, your principles, your voice. The people you trust ask it. It answers with a confidence it is willing to show. You approve, correct, or disavow, every time.
One clarifying question at most. Then it reasons against your past decisions — never improvises.
Every answer lands in your inbox first. Approve, correct, or disavow — a correction reaches the asker in seconds.
Nothing moves without your yes.
The queue forms quietly. A vendor advance waits on one nod. A facade specification waits four days for an answer the Managing Director could give in ninety seconds. Nobody in that queue is short of information — they are short of him. The cost was never the decision. It is the wait around it.
Mini-me does not stand in for that judgment. It answers from the record of it — his own past decisions, in his own words — and every answer reaches him before the person who asked can act on it.
THE DESIGN GOAL — TIME FOR A DECISION TO REACH THE PERSON WHO NEEDS IT
TRAINED ON A YEAR OF HIS OWN SENT MAIL AND EVERY TRANSCRIBED MEETING
THE HARD CAP ON HOW LONG THE AGENT MAY REASON BEFORE IT ANSWERS
Every ask runs the same five steps — one Claude agent loop with MCP-shaped tools, grounded in his own record, capped at six tool-call rounds, and streamed to the asker as it works. Bounded, auditable, never improvised.
Someone he has granted access to describes a real situation, in text. One ask, one answer — there is no open chat window to wander into.
If the ask is ambiguous, the twin returns one bounded multiple-choice question — never an open-ended back-and-forth. Pick an option and the loop continues.
It embeds the ask with voyage-3-large at 1024 dimensions, pulls the nearest prior decisions from his corpus with pgvector, reranks top-8 down to top-5, and reasons against his stated principles — his record only, never another twin's.
A directive or a piece of advice, in his voice, carrying a confidence score and the exact past decisions it rests on. Below its own threshold it escalates to the strongest model — or it declines and says: ask him directly.
It reaches his inbox before the asker can act on it. Approve, correct, or disavow. A correction reaches the asker in seconds — and goes back into the corpus the twin reasons from.
The same five steps, on the screens the pilot actually uses.
It knows when it is not sure.
Built on Claude, with the reasoning explicitly rationed — a cheap model for the small steps, a strong model for the answer that carries his name, and a ceiling on how long any single ask may run. Everything below is one request path.
Expo · iOS + Android
text only, by choice
Cloud Run · asia-south1
Mumbai
one loop · one mind
≤6 tool-call rounds · prompt caching
find_similar_decisions
search_corpus
get_recent_meetings
search_outlook_mail
Cloud SQL · Mumbai
row-level isolation
top-8 → rerank → top-5
Fireflies API key
Microsoft 365 OAuth
no server-wide identity
tool-call events streamed to the client
approve · correct · disavow
a correction reaches the asker in seconds
A light model handles the clarifying question, the sorting and the rerank. The committed answer — the one that carries his name — is written by a stronger model. The strongest is held in reserve and called only when confidence falls below threshold.
haiku → clarify · sort · rerankEvery ask is embedded and matched against the owner's corpus of past decisions, meetings and mail — never another twin's. The nearest matches are reranked before they reach the model that drafts the answer.
voyage-3-large · 1024-dimA hard ceiling of six tool-call rounds per ask. If the answer is not there by then, the twin says so rather than continuing to guess — a limit chosen on purpose, not a limitation discovered by accident.
max 6 tool-call roundsA score never travels alone. Every confidence number is shown with the past decisions it rests on — dated, quoted, and scored by how close they sit to the question that was asked.
Illustrative. The rows are always the owner's own decisions, retrieved from his own corpus — never another twin's.
When the record does not support an answer, the twin does not round up. The ask is pinned to his inbox, escalated to the strongest model, or declined outright with the plain instruction to ask him directly.
It knows when it is not sure.
A directive is a call he would make today. Advice is the twin telling the asker to treat it as a starting point — and to go and find him.
A twin is not built from documents about the company. It is built from the moments its owner actually made a call — what he chose, what he refused, and the reason he gave at the time.
The rooms he was actually in. Fireflies transcribes them; the twin keeps the calls that were made and the reasoning that carried them, not the small talk around them.
Three hundred and sixty-five days of what he wrote back — the short replies where most decisions in a company are actually taken.
When he corrects an answer, the correction does not only reach the person who asked. It returns to the corpus, and the twin is nearer to him than it was that morning.
Transcripts and sent mail are read through the owner's own connectors, on his schedule, into his own store.
fireflies · microsoft 365Quoted threads and signatures are stripped so the model learns from what he wrote, not what he was replying to. Each passage is then typed.
decision / principle / preferenceEach passage is embedded and stored against his twin. One corpus per owner, isolated at the database row.
voyage-3-large · 1024-dim
‘Your presence can be everywhere. And you can be everywhere.’ROHIT GERA · MANAGING DIRECTOR, GERA DEVELOPMENTS
Access to a twin is granted by its owner, one person at a time, and it points in one direction: letting someone consult your twin does not let you consult theirs. That is a separate request, the other way. Revoke a grant in a single gesture and the twin stops answering for that person.
One corpus per twin, held in Postgres with pgvector on Cloud SQL in Mumbai. Every retrieval is scoped to a single owner’s rows before it runs, so a twin cannot reach a corpus that is not its own. Accounts are admin-provisioned; there is no self-signup.
Meetings and mail connect through the owner’s own credentials — a personal Fireflies API key, a Microsoft 365 OAuth token — stored against that one user. There is no server-wide identity for either provider, so nothing keeps reading when a key is withdrawn.
He decides who is in the room. The twin never opens the door.
One app, both platforms, built in Expo. It opens to whoever you are — your own twin on one side of the day, the people who trust you with theirs a tap away.
An asker describes the real situation in their own words. One bounded clarifying question at most — never an open chat window.
A directive or advice, depending on who is asking, carrying the confidence it is willing to show and the past decisions it reasoned from.
A notification the moment it is given. Approve, correct, or disavow — a correction lands with the asker in seconds.
The same four come up in every first conversation. None of them has a soft answer — each one is settled by how the system is built.
It is labelled as his twin, never as him. Every answer arrives with its confidence and the past decisions it drew on, and he sees it before the asker can act. Text only, by choice — no face, no voice clone.
answer · confidence · cited decisionsHe is. Every answer lands in his inbox first: approve, correct, or disavow. A correction reaches the asker in seconds and retrains the twin; a disavowed answer is retracted to the person who asked it.
owner notified before the asker can actIt shows its work, so a wrong answer is visible rather than plausible. Below its own confidence threshold it escalates to the strongest model, pins the ask to his inbox, or declines outright and says: ask him directly.
low confidence → escalate · pin · declineIn Mumbai. Postgres with pgvector on Cloud SQL, behind FastAPI on Cloud Run in asia-south1. One corpus per twin, isolated at the row; connectors run on your own keys. Nothing trains a shared model, and access is revocable in one gesture.
asia-south1 · row-level isolation · per-user keysA fifth question? Ask a person — hello@askminime.ai
Mini-me is in a private pilot at Gera Developments — one trained twin, three people trusted to consult it. Seats open by invitation. Tell us whose judgment your company can least afford to wait for.
We read every request personally. When a pilot seat opens for your company, the first email you get will be from a human, not a twin.