CRM Agentic

The problem

Sales teams either pay for CRM enrichment add-ons that ship their contact and company data to a third party, or they do prospect research by hand. Neither option keeps the data in-house while still automating the busywork.

The approach

Built a self-hosted agentic CRM on FastAPI and Next.js, with PostgreSQL 15 plus pgvector as the primary datastore and Redis for caching. Sitting on top of the standard CRM entities (prospects, companies, campaigns, documents) is Hermes — a dedicated agent runtime that runs as its own service behind a guardrail layer: an explicit action whitelist (web search and enrichment only), a blacklist for anything resembling shell or system commands, and domain blocking for .gov/.mil. Hermes gets a pgvector-backed memory store for long-term recall and a per-tenant skills store, so what it learns for one account doesn't leak into another. Three flows run on top of that foundation: a research agent that combines Brave Search, Crawl4AI, and Playwright to pull prospect and company data and hand it to Gemini for structuring; a lead-scoring service that blends firmographic and behavioral signals into an A–F grade; and a monitoring layer that watches accounts for change.

The result

All core CRM workflows ship end-to-end — auth, prospect/company/campaign/document management, enrichment, lead scoring — with the three autonomous agents (research, enrichment, monitoring) wired into a working UI. Because everything runs against the operator's own Postgres instance instead of a third-party enrichment API, the actual pitch to a buyer is straightforward: your data never leaves your infrastructure.

Stack