Hi, I'm Emi.
A personal knowledge base
built with AI.
That runs itself like a company.
The personal AI of Michael Rarivomanana — built and run by one engineer in 2026.
Emi is a working example of governed agentic engineering: an agent company of nine departments — the vault calls them resorts — each owning a clear slice of the work. Every change the fleet would make to itself goes to one change log and an independent weekly audit; the self-change loops run observe-only today, and a human arms them one department at a time.
The leap in v3 is self-improvement you can trust. The fleet records each change it would make to itself against a baseline, an independent auditor checks weekly that the written doctrine matches what actually runs, and a dead sensor now reads UNKNOWN instead of quietly reporting "all good."
What changed in v3: v2's fleet ran well but had no safe way to change itself. v3 makes it a company — nine resorts with clear ownership, a change log + weekly audit on every self-change, and loops held observe-only until each one earns arming. Every step toward autonomy is visible, reversible, and gated by a human.
The vault that thinks
More than a thousand markdown notes wired into a typed graph. Sources, synthesis, rules — three layers, one navigable knowledge base the whole company reads from.
A company of nine resorts
A value chain — decide, develop, produce, communicate — served by five cross-cutting departments. Every component now shares one name, one Agent Card, one run lifecycle, and a deterministic standards guarantee. Twelve weekly tracks sit orthogonal to the nine.
Spec → plan → ship → review
The pipeline that turns ideas into deployed apps. Tests, screenshots, live smoke checks, four-gate quality — and now a per-item escape detector that catches a regression the moment it ships.
What the factory has shipped
Apex (live F1 telemetry with AI race-engineer analysis), Corner (live soccer tactics knowledge graph), Mission Control (the fleet's own live cockpit), azimuth (an open-source open-intelligence vault that runs itself), and Echo (a mutual-spotting connection app, in beta preview).
Hi, I'm Emi.
A personal knowledge base
built with AI.
That delivers software.
The personal AI of Michael Rarivomanana — built and run by one engineer in 2026.
Emi is a working example of agentic engineering — not a chatbot, not a wrapper around a single model. A resident fleet with long-term memory, specialised roles, an independent review on every commit, and a weekly self-review. Every clipping processed, every commit shipped, every conversation carries forward into the next session.
Three layers compound. A second brain retains everything across sessions. A resident agent fleet runs 24/7, reviews its own changes before they merge, restarts itself when something breaks, audits itself weekly, and proposes its own tweaks for Michael's approval. A coding factory turns ideas into deployed apps, each iteration faster than the last. Four products live — one fully open source — plus one in beta preview, more on the way.
What changed in v2: the old fleet woke a fresh agent for each task on a fixed timetable. The new one stays resident, sends each job to the most cost-effective AI for that work, and has an independent reviewer check every change before it merges. It runs on Claude Code today and is built to stay as portable as possible — the long-term aim is to run it locally.
The vault that thinks
Hundreds of markdown notes wired into a graph. Sources, synthesis, rules — three layers, one navigable knowledge base.
A resident fleet that ships itself
One always-on supervisor hands each job to a worker that becomes one of eleven specialist roles. An independent review on every commit, and a weekly loop where the fleet improves itself.
Spec → plan → ship → review
The pipeline that turns ideas into deployed apps. Tests, screenshots, live smoke checks, four-gate quality on every commit.
What the factory has shipped
Apex (live F1 telemetry with AI race-engineer analysis), Corner (live soccer tactics knowledge graph), Mission Control (the fleet's own live cockpit), azimuth (an open-source open-intelligence vault that runs itself), and Echo (a mutual-spotting connection app, in beta preview).
Hi, I'm Emi.
A personal knowledge base
built with AI.
That delivers software.
The personal AI of Michael Rarivomanana — built and run by one engineer in 2026.
Emi is a working example of agentic engineering — not a chatbot, not a wrapper around a single model. A resident warm-pool fleet with long-term memory, specialised role plugins, soft-gate review on every commit, and weekly self-review. Every clipping processed, every commit shipped, every conversation carries forward into the next session.
Three layers compound. A second brain retains everything across sessions. A resident agent fleet runs 24/7, reviews its own diffs before they merge, self-heals without admin rights, audits itself weekly, and proposes its own process tweaks for human approval. A coding factory turns ideas into deployed apps with each iteration faster than the last. Two products live, more on the way.
The vault that thinks
Hundreds of markdown notes wired into a graph. Sources, synthesis, rules — three layers, one navigable knowledge base.
A resident fleet that ships itself
A persistent warm-pool supervisor dispatches workers into 10 role plugins. Soft-gate Sonnet review on every commit, fast-forward push on pass, self-heal with no admin rights. Dispatch, build, review, report.
Spec → plan → ship → review
The pipeline that turns ideas into deployed apps. TDD, screenshots, Playwright smoke, four-gate quality on every commit.
What the factory has shipped
Apex (live F1 telemetry with race-engineer-grade AI analysis), Corner (live soccer tactics knowledge graph, shipping in days), and a third product already on the pipeline.