autoRFP

Inspiration

RFP sites are fragmented and inconsistent. Teams waste hours finding relevant opportunities, parsing messy PDFs, and stitching answers from past docs—only to send dozens of proposals before winning one.

What it does

autoRFP is an end-to-end agent that monitors public RFP portals, cleans and normalizes PDFs into a schema-safe context pack, aligns each requirement to your company’s evidence, drafts citation-backed answers, self-evaluates coverage, and publishes a ready-to-send proposal with a 1-page microsite.

How we built it

Bright Data (MCP) discovers live RFPs and links → Apify extracts PDF/text → Senso converts raw text into a structured RFP schema (requirements, scoring, deadlines) → Redis VL + LlamaIndex build the retrieval graph over your docs + RFP → the agent drafts answers with inline citations → HoneyHive evaluates coverage and hallucination risk and triggers targeted rewrites (A2A loop) → Stytch gates a preview → Qodo publishes a shareable proposal page.

Challenges we ran into

Credential and infra setup across multiple sponsors, rate-limits on web access, and getting reliable PDF structure. Prompting for strict citations without bloating responses was tricky. Mapping heterogeneous RFP formats into one schema and keeping IDs stable for iterative rewrites also required care.

Accomplishments that we’re proud of

A clean, reproducible context pipeline that goes from a live RFP URL to a citation-backed proposal in minutes. The self-repair loop measurably increases requirement coverage, and the one-click publish makes the demo tangible for judges and real users.

What we learned

Good “context engineering” beats bigger context windows: a tight schema and deterministic IDs enable targeted regeneration. MCP web access plus an extraction actor is enough for dependable ingestion if you normalize early. Redis VL made retrieval fast and debuggable. Automated eval (coverage + citation checks) is the fastest way to reduce hallucinations under time pressure. Clear separation between ingest → normalize → retrieve → draft → evaluate → publish keeps the system resilient.

What’s next for autoRFP

Focus on one vertical (e.g., GovTech IT, healthcare, or security questionnaires) to ship a specialized schema and templates. Add vendor-security and grant-proposal modes, human-in-the-loop redlining, and team workflows. Package as a “RFP-in, proposal-out” SaaS with audit logs and privacy controls.

Built With

• apify
• brightdata
• cursor
• honeyhive
• llamaindex
• mcp
• python
• qodo
• redis
• redisvl
• senso
• stytch
• typescript