GrantMatch — AI Grant Discovery & Proposals

Finding the right grants is slow and manual. A business has to comb through hundreds of funders, work out which programmes it is actually eligible for, track deadlines, and then write a tailored proposal for each one. Plain keyword search returns too much noise and still misses grants that are semantically relevant, and eligibility often hinges on details buried in funder pages. The client needed this to run continuously and at scale, not as a one-off search.

Automate the full grant lifecycle for a business: build an accurate profile, continuously discover new and recurring grants, match the business to the grants it can realistically win, and help draft the proposals, with matches that are relevant and defensible rather than a wall of keyword hits.

A profile is built by crawling the business site and using an LLM to classify pages and pull out what matters for eligibility. Grants are discovered by scraping funder sites with Playwright and extracting structured fields with an LLM, including recurring grants. Matching runs as hybrid retrieval on MongoDB Atlas: a vector search over embeddings and a full-text search run in parallel, their results are combined with Reciprocal Rank Fusion, a reranker re-scores the top candidates, and an LLM fact-checks each match against the profile so weak or ineligible grants drop out. A LangGraph agent sits on top, routing across OpenAI and Anthropic models and calling tools (text generation, image analysis and generation, and platform APIs) to draft and refine proposal content. Everything is wired together as FastAPI microservices communicating over AWS SQS.

1. 1Crawl the business website; an LLM classifies pages and extracts the profile
2. 2Scrape funder sites with Playwright; an LLM extracts structured grant data
3. 3Embed grants and run vector plus full-text search on MongoDB Atlas
4. 4Fuse the two result sets with Reciprocal Rank Fusion, then rerank the top candidates
5. 5An LLM fact-checks each grant against the profile to drop weak matches
6. 6A LangGraph agent drafts and edits proposal content, calling tools and platform APIs

• Hybrid retrieval: MongoDB Atlas vector search + full-text search
• Reciprocal Rank Fusion to combine the two rankings
• Reranking and LLM fact-checking of matched grants
• LLM page classification and structured grant extraction
• LangGraph agent with multi-model routing (OpenAI + Anthropic) and tools
• Playwright scraping and PyMuPDF / OCR document processing

• End-to-end: profile, discover, match and write
• Semantic and keyword matching, not just keywords
• Continuous discovery of new and recurring grants
• Eligibility-aware matches, fact-checked by an LLM
• Agent-driven proposal drafting and editing
• Event-driven microservices that scale independently

The platform is a set of FastAPI microservices, each owning a domain (profile building, grant discovery, matching, curation and content generation) and communicating over AWS SQS. Grants and embeddings live in MongoDB Atlas, which provides both vector and full-text search. Retrieval combines the two with Reciprocal Rank Fusion, followed by reranking and an LLM fact-check. A separate agent service, built on LangGraph, exposes AI capabilities behind an aggregator API and routes each task to the right model and tool. Services are containerized with Docker.

GrantMatch is delivered as an API-first platform: the services expose endpoints for profiling, discovery, matching and content generation, and emit events over SQS, so a product UI can surface matched grants, deadlines and draft proposals per business without the matching logic living in the front end.

• MongoDB Atlas (vector + full-text search)
• OpenAI and Anthropic (via LangChain / LangGraph)
• AWS SQS (event-driven microservices) + boto3
• Playwright and BeautifulSoup (scraping), PyMuPDF / pdf2image / OCR

Each service is a Dockerized FastAPI app, deployed independently and talking over SQS, which lets discovery, matching and generation scale on their own. MongoDB Atlas handles storage and search, and the agent runs as its own service behind an aggregator API.