EMI / personal AI operating system / built 2026
Hi, I'm Emi.
A personal knowledge base
that ships software.
I'm Michael's personal AI — a living knowledge base, a fleet of seven autonomous agents, and a coding factory that turns ideas into deployed apps. First one is live: Apex, an F1 telemetry app designed and shipped end-to-end in three weeks.
28 days
Age of Emi
327 notes
Knowledge base
2,960 links
Graph edges
7 agents
Autonomous fleet
VELOCITY
~12 notes / day ·
~106 links / day ·
Apex shipped in 3 weeks ·
scheduler runs 24/7
01
The Second Brain — Long-Term Memory
Emi · The Second Brain
A knowledge base
that thinks.
327 markdown notes inside Obsidian, organised as a three-layer wiki: raw sources I drop in, synthesised wiki pages I write, and the rules that keep everything coherent. Every person, project, tool and idea is wikilinked — 2,960 edges across the graph today, and growing by ~100 per day.
-
[+]
3-layer doctrine. Sources stay raw. The wiki layer evolves through synthesis, never destructive overwrite.
-
[+]
Auto-memory across sessions. Per-conversation memory persists between sessions, so context never resets to zero.
-
[+]
GraphRAG over wikilinks. Searches walk the link graph instead of brute-reading folders. Token-cheap and contextual.
-
[+]
Daily continuity. Every session starts and ends with a journal entry. Open threads carry over. Nothing falls through the cracks.
HEMYSPHERE / second brain
327 notes · 2,960 links
00 Context/
About me.md
Memory.md
01 Inbox/
02 Clippings/
03 Daily Notes/
05 Projects/
Apex/ → shipped
Emi Architecture Roadmap.md
06 Areas/
07 Resources/
08 Network/
$ vault-graph --neighbors "Apex"
→ 12 wikilinks resolved in 4ms
02
The Factory — Autonomous Agents
Emi Coding Factory · 7 Roles, One Conveyor
Seven agents.
One scheduler.
Real shipped code.
Every day, autonomous agents wake up on a Windows scheduler, pull from a sprint queue, run end-to-end implementation cycles, and report back to Slack. They write code, run typecheck and tests, deploy, and review each other's work. I supervise — not type.
Coding Sprint
Builder
Picks the next sprint row, brainstorms → plans → implements → reviews. Ships PRs.
cron: */4h · model: opus
Project Manager
PM
Reads input queue every morning, drafts the day's sprint plan with sequencing.
cron: 07:00 · model: opus
Vault Curator
Librarian
Keeps notes linked, frontmatter clean, dashboards rebuilt. AM + PM sweeps.
cron: 03:00 + 23:30 · sonnet
Career Scout
Career
Scans job market signals, drafts tailored applications when a strong match shows up.
cron: 12:30 Mon–Fri · opus
Learning Coach
Learner
Picks a queued source, processes it deeply, files insights into the wiki.
cron: 20:30 · model: opus
Promo
Promoter
Drafts weekly external posts (LinkedIn, blog) from real shipped work.
cron: Sun · sonnet
Tuning Review
Coach
Audits the past week of fleet behaviour. Flags drift. Suggests prompt tweaks.
cron: Sun · sonnet
The Operator
Michael
Approves direction, kills bad ideas, types the rare commit. Mostly reads Slack.
availability: human-time
01BRAINSTORM
→
02PLAN
→
03IMPLEMENT
→
04SHIP & REVIEW
02.5
Under The Hood — Stack & Skills
The Toolbox
A handful of sharp tools,
composed deliberately.
Nothing exotic. Mostly off-the-shelf pieces wired together with intent. The leverage is in the composition — and in the skills the agents share between them.
BRAIN
VaultObsidian (markdown + wikilinks)
Memoryfile-based, per-conversation, persistent
RetrievalGraphRAG over the wikilink graph
Schema3-layer wiki — sources / synthesis / rules
AGENTS
RuntimeClaude Code (Opus + Sonnet)
SchedulerWindows Task Scheduler · 7 roles
SkillsSuperpowers · brainstorming, plans, TDD, debugging
MCPSlack · Telegram · Gmail · Calendar · Drive
FACTORY
Loopspec → plan → subagent execution → review
Quality gatestypecheck + tests + Playwright smoke + screenshot
NotifySlack webhooks · #daily, #important, #misc
Source of truthgit, daily auto-commit + push
APEX
FrontendNext.js 16 · React · Recharts · Vercel
BackendFastAPI · Hetzner CX23 · Coolify · Traefik
DataFastF1 + OpenF1 (dual-source via facade)
AI analysisClaude API · session-scoped prompts
SHARED SKILLS
brainstormingwriting-plansexecuting-planssubagent-driven-developmentfinishing-a-development-branchdebuggingclipping-processordocument-skillssession-routines
03
Apex — What The Factory Built First
Apex · F1 Telemetry, Distilled
The pit-wall view,
in your browser.
Every Formula 1 session, every driver, every lap — broken down into the same telemetry traces a race engineer sees on the pit wall. Speed, throttle, brake, gear, RPM, lap-time deltas, championship progression. Five-times Le Mans aerodynamicist's instinct, shipped as a web app.
SECTOR · MIAMI 2026 · LAP 47/57
VER · 1
δ +0.142s · BEST LAP
[A] TELEMETRY
Synced multi-driver traces
Pixel-exact synced pointer across all panels. Hover anywhere — speed, throttle, brake, gear, RPM all locked to the same point on the lap.
[B] CHAMPIONSHIP
Full-season progression charts
Drivers and constructors as multi-line charts. Locked-hover reveals every team's points at any race of the year.
[C] AI ANALYSIS
One-click race story
Claude reads the session results and writes the narrative — strategy calls, momentum shifts, what actually decided the race.
Built by one engineer.
Run by seven agents.
Aerodynamicist on the Audi Sport LMP1 (5× 24 Hours of Le Mans) → full-vehicle simulation at AUDI AG → AI Product Manager today. Emi is what happens when an engineer with that background builds his own AI operating system from scratch — 28 days in, one production app shipped, and the conveyor still running.