Back to Portfolio
ActiveNov 2025 – Present

amoOS: My Personal Software Factory

An AI operating system where I plan, dispatch, and verify while a fleet of AI agents does the building. It runs every project I own, remembers everything it learns, and reports to me through a native macOS cockpit and Telegram.

11
Projects Orchestrated
400+
Compiled Knowledge Pages
3
Command Surfaces
4
Factory KPIs Tracked

Architecture

  • A fleet of AI coding agents executes the actual work across all of my projects. My role is reduced to planning, dispatching, and verifying
  • Everything the system touches (sessions, decisions, documents, forwarded links) is distilled into a living per-project wiki plus a personal knowledge base, so context compounds instead of evaporating
  • Multi-surface command: a native macOS cockpit for deep work, Telegram for mobile steering and approvals, and a web dashboard for reference
  • Frictionless capture: forward a link or a note to Telegram and it lands in the knowledge base, enriched and searchable
  • A shared project registry, sprints, kanban, and cross-project memory search that every work session plugs into
  • Self-accounting: the factory measures its own utilization, yield, attention cost, and reporting speed, and answers for them

Key Decisions

Human verification gates everything the fleet ships

Why: Autonomous output without trust is worthless. 'Verified done, with evidence' is the unit of progress

Tradeoff: Throughput is capped by my review attention, which forced me to design the review experience as a first-class product

One system runs every project instead of per-project tooling

Why: Memory compounds across projects: a lesson learned in one product immediately benefits the next

Tradeoff: Platform work permanently competes with product work, so scope discipline is enforced ruthlessly

The system owns the memory, not the humans or the chat logs

Why: An assistant that forgets is a chatbot; a system that compounds knowledge is a moat

Tradeoff: Knowledge upkeep is real work: curation, deduplication, and staleness detection all had to be built

Technologies

AI Agent FleetOrchestrationSecond BrainMCPFastAPINext.js

What I Learned

  • Once agents write the code, the bottleneck moves to deciding, verifying, and remembering, so that is what the system optimizes for.
  • AI agents are only as good as the context you hand them; a compiled, current knowledge base beats raw history every time.
  • Trust is the real product: slower verified output is worth more than fast unverified output.