Sourcery: A Personal AI Infrastructure Architecture

See the personal AI system that reads Paul's world and acts on it.

An interactive walkthrough of Sourcery — the local-first AI infrastructure Paul runs on a Mac Mini. Fourteen external services flow through a prompt-injection gate into a Markdown corpus that becomes both a knowledge graph and a hybrid search index, which roughly fourteen agents reason over and act on.

What you'll get

• One Markdown corpus as the single source of truth: fourteen connectors — Gmail, Slack, Drive, ClickUp, HubSpot, Granola, iMessage, and more — each write clean Markdown into sources/, and nothing downstream is allowed to invent facts.
• A prompt-injection gate on everything inbound: pi_scan scores each item with delimiter, encoding, multilingual, and unicode-obfuscation detectors, holds high-risk items for a human Save / Reject on Telegram, and audits every verdict in Supabase.
• A knowledge graph and a hybrid index, both derived views: Graphify extracts entities and relationships into graph.json while a local fastembed pipeline builds a 384-dimension vector index — every node points back to its source file.
• Retrieval that fuses three inputs into one ranked answer: structured entity keys, full-text search, and vector similarity fused by RRF, reachable through graph-traversal, hybrid-search, and raw-SQL MCP surfaces.
• Agents that reason and act, with a human in the loop: ~14 agents recommend and draft; a Telegram approval flow keeps Paul in control before anything irreversible, all scheduled by ~40 launchd jobs on the Mac Mini and mirrored to the MacBook over Tailscale.