Wecelium

Wecelium was a self-directed proposal for an institutional knowledge system for AI-assisted software engineering. The system solves the problem of scattered institutional knowledge by capturing architectural decisions, technical debt, conventions, and code insights during normal engineering practice, then providing persistent, contextual knowledge to developers and LLMs alike. The architecture is Postgres-native, using a three-table-family schema that embodies the system's methodology: personas and problems as rows plus join tables (not columns/tables), allowing traceability via database queries and reverting migrations via Sqitch. The system maintained over 14,000 knowledge entries across 11 internal projects. Search is the core differentiator. Wecelium implements a three-signal Reciprocal Rank Fusion (RRF) that runs three retrieval strategies in parallel and fuses results without manual weight tuning: 1. **Vector similarity** — pgvector (768-dim, nomic-embed-text via Ollama) captures semantic intent 2. **Full-text search** — GIN indexes capture keyword/structural matches 3. **Graph traversal** — recursive CTEs traverse the knowledge graph to find structurally related items The system connects related knowledge across dimensions: decisions → constraints → trade-offs → tech debt. A 1,352-item pending review queue with staging confirm gates captures knowledge organically during active coding sessions before it enters the live database. The interface consists of 30+ MCP tools for knowledge interaction, a PostgREST API auto-generated from the schema, and a session hooks system for session capture and progressive context loading. Status: The proposal was submitted as a pitch for adoption. The business case was not accepted and the system was decommissioned. The technical architecture, schema design, and search implementation remain as a reference of engineering work.