A production knowledge brain for every client.
Architecture of the Company Brain: five knowledge sources (Slack, Drive, Gmail, Notion and GitHub) feed an ingest pipeline, then hybrid BM25 plus vector retrieval with RRF fusion, exposed through an MCP server with 25+ tools that powers agents and client dashboards.
mcp: FastMCP server, 25+ tools, Cognito auth, per-client row-level security
- Problem
- Amplifyr helps brands win AI search: measuring how ChatGPT, Perplexity and Google AI Overviews cite and recommend them, then optimizing that visibility (GEO). Inside that mission, client knowledge sat scattered across Slack, Drive, Gmail, Notion and GitHub: unsearchable, unauditable, unusable by agents.
- Built
- The Company Brain: a custom MCP server (FastMCP, 25+ tools) doing hybrid BM25 + vector retrieval with RRF fusion over Aurora Postgres/pgvector and Bedrock Titan embeddings, deployed on Fargate with Cognito auth and per-client row-level security. Plus multi-stage LangGraph content agents with automated fact-checking, and 15 client-facing analytics visualisations in Next.js.
- Trade-offs
- Hybrid retrieval over pure vector: BM25 catches the exact-term queries embeddings miss, RRF fuses without tuning a weighted sum. Deterministic record/replay harness over live-LLM tests, so sub-agent regressions surface in CI, not in front of clients.
- Outcome
- In production as internal infrastructure: the whole team and its AI agents run on it daily, boosting operational efficiency across every client engagement, with retrieval-quality and claim-fidelity eval gates wired into CI.
