Procurion - Autonomous Procurement Intelligence
A multi-agent procurement system that ingests an enterprise RFP, scouts vendors, verifies every requirement against the document with retrieval-augmented evidence, and drafts grounded negotiation emails in a single halting LangGraph run.
- Python
- LangGraph
- FAISS
- FastAPI
- Node.js/Express
SYSTEM BUILD
What went into it
- 01
Built a three-agent LangGraph pipeline — scout, compliance, negotiation — fanned out per vendor via the Send API with deterministic state reducers, invoked exactly once per run.
- 02
Grounded every compliance verdict and discount ask in a per-run FAISS index built with heading-aware chunking, citing specific chunk IDs in the final report.
- 03
Designed an LLM facade with automatic Gemini → Groq → Anthropic/OpenAI failover that degrades to a deterministic engine and records exactly which tasks degraded.
- 04
Streamed live agent progress to a React console over Server-Sent Events with Last-Event-ID replay, backed by a Node gateway and SQLite run store.
MEASURED OUTPUT
What changed
Compressed vendor discovery, requirement verification, and outreach preparation into one auditable run — every verdict cited to a chunk of the client's own RFP and the system halts when the report lands.
Shipped with 68 green tests across pytest, node:test, and vitest suites plus a Playwright browser E2E that uploads a real PDF and waits for the report.