AI systems
Personal AI Ops
Private repository — the publicly agreed part is shownFrom chat to tasks: a conversation becomes a project with a to-do list
Problem
Project agreements live in Telegram chats and get lost in the scroll. Nobody moves them into a tracker by hand, and a fully autonomous agent that writes to people on its own is a risk I won't take.
System
A fixed six-step pipeline: Telethon reads only chats on an explicit whitelist, the snapshot is normalised, Gemini with a strict response_schema parses it into a project and tasks, Pydantic validates the reply, the result goes to Postgres, then Notion, then a notification arrives from a separate bot. The boundaries are deliberately tight: read, analyse, structure, save, notify. No outbound messages to people, no agent swarms. OCR and transcription are designed as future services that plug in without touching the orchestrator.
Role
Solo build: architecture, data model, integrations.
Outcome
A working MVP with a clear limit on authority: the system cannot message anyone behind my back. The orchestrator stays linear and readable; extensions go in as services, not as conditionals inside it.