A working rebuild of a compound AI grant-discovery agent: one unified index of federal and foundation opportunities, searched by an agent that reads your proposal and shows its reasoning as it goes.
Open the app How it worksDrop in a proposal PDF or just type a sentence about your research. The agent extracts the terms itself, queries a unified index of roughly 12,000 opportunities, and only reaches for live web search when something might have posted in the last few days. Every result links straight to the official agency or funder page. It will not invent a program, a deadline, or a link.
Grants.gov, live: NSF, NIH, DARPA, DOE, USDA and the rest, normalized into one schema.
Private foundation programs via Kindora, the largest slice of the index.
Upload a draft and the agent works from your actual methods and goals, no keyword guessing.
Tool calls and reasoning stream in real time so you can judge why a result surfaced.
The design follows Tang and Kejriwal's "A Compound AI Agent for Conversational Grant Discovery" (USC/GRAIL, 2026): an aggregation layer that scrapes and normalizes opportunities, and a ReAct agent with two tools, a structured index search and a live web search. Their deployment at grail.page reports 3,000-plus users. This is an independent build of the same architecture, with Kindora standing in for the foundation-scraping half.
Their paper reports discovery time falling from 30 to 45 minutes of manual portal searching to under 10 minutes. Worth being precise about that number: it is a self-reported deployment observation, not a controlled benchmark. The authors say so plainly, writing that they offer no formal evaluation. I quote it as a design target, not a proven result. More on what is and isn't measured.