GrantRadar

Turn any grant announcement, RFP, or funding notice into clean, structured JSON — from a URL, a PDF, or raw pasted text.

GrantRadar reads government grant listings, agency funding notices, and RFP documents and extracts the fields that actually matter: funder, award range, deadlines, eligibility, opportunity number, and contact info. Point it at a grants.gov or agency page, a direct PDF link, or paste the text yourself. Extraction is powered by a DeepSeek LLM and results are normalized into a consistent schema with ISO dates and integer dollar amounts.

Endpoints

Endpoint What it returns

POST /extract A structured grant record extracted from a url, pdf_url, or raw text. Wraps the result as { "grant": {...}, "cached": <bool> }. The grant object includes: title, agency, program, opportunity_number, funding_type (one of grant, cooperative_agreement, contract, fellowship, prize, loan, other), award_floor, award_ceiling (integer dollars), eligible_applicants (array), cost_share_required (bool), key_dates (posted, application_due, award_notification as ISO dates), categories (array of topic tags), application_url, contact (name, email, phone), summary, and source_url.

GET /health Liveness check — returns { "status": "ok" }.

Request body for /extract (provide at least one of):

url — a web page to fetch and parse (grants.gov, beta.sam.gov, agency site, etc.)
pdf_url — a direct link to a PDF announcement
text — raw grant announcement text you already have

Why this API

One schema for everything. Whether the source is HTML, a PDF, or pasted text, you get the same normalized GrantRecord — dates as YYYY-MM-DD, money as plain integers, no $ or commas to clean up.
Built for messy government documents. The extractor is specifically tuned to recognize opportunity-number formats (NOFO, FOA, CFDA, solicitation, announcement numbers like DE-FOA-0003456 or HHS-2025-ACF-OHSEPR-NR-0124) and to surface the real application URL, even from beta.sam.gov and grants.gov package pages.
Three input modes, zero scraping infrastructure on your side. GrantRadar fetches and cleans the page (stripping nav/scripts/boilerplate) or extracts PDF text for you via PyMuPDF — you just pass a link.
Cached by request. Identical requests are served from cache (keyed by SHA-256 of the input), so repeat lookups are fast and don't re-bill the LLM.
LLM-normalized, not regex-fragile. Eligibility, categories, and funding type are inferred into consistent tags rather than copied verbatim.

Typical use cases

Build a grants aggregator or newsletter that ingests agency listing pages and outputs uniform records.
Auto-populate a CRM or tracking sheet with deadlines, award ceilings, and contacts pulled from RFP PDFs.
Screen inbound funding notices for eligibility (eligible_applicants, cost_share_required) before a human reviews them.
Normalize a backlog of saved grant pages or PDFs into a searchable, structured dataset.
Power a Slack/email bot that turns a pasted funding announcement into a clean summary card.

Good to know

At least one input is required. A request with none of url, text, or pdf_url returns 422. Empty fetched content also returns 422.
All grant fields are nullable. Any field the document doesn't contain (or that can't be inferred) comes back as null; eligible_applicants and categories default to [] and key_dates/contact to objects of nulls. Treat every field as optional.
Caching is 24 hours by default. Responses are cached per unique request payload (TTL 86400s); a cache hit sets "cached": true. If a page updates within that window you may receive the prior extraction.
Fetched content is truncated. Pages and PDFs are capped at ~40,000 characters before extraction, so very long documents are processed only up to that limit.
Fetch and extraction failures surface as 502. A URL that can't be retrieved, a PDF that won't parse, or an LLM/validation error returns 502 with a detail message.
Extraction quality depends on the source. Output is LLM-inferred from whatever text the document provides; values like funding_type, categories, and inferred dollar amounts should be reviewed for high-stakes decisions. This is an extraction tool, not an official system of record.