Skip to content
fetching.[)esign_
AI / Full Stack

Open Brain (OB1)

Self-hosted knowledge system with pgvector semantic search and an MCP server.

DenoHonoPostgreSQL + pgvectorMCPNext.jsOpenAI embeddings

Problem

No unified, self-hosted, semantically searchable store for thoughts, contacts, and work that also handles dedup and secret-safety.

What I built

A production knowledge system: pgvector + embeddings (1536-dim, HNSW) for sub-100ms retrieval, an MCP server exposing ~12 REST and 12 MCP tools, five ingestion sources with automatic metadata extraction and dedup, and a dashboard for timeline, stats, contacts, and job pipeline.

Engineering highlights

Cosine similarity over HNSW; SHA256 content-hash dedup with upsert returning duplicate/orphan status; LLM-powered metadata extraction; a multi-table schema with expression indexes; regex secret detection to block credential storage; a dual REST + MCP API on Hono.