Opportunity Radar

The chore nobody has time for

Every year, billions in public funding goes unclaimed. Not because people don't qualify — because finding out whether you qualify means reading a government database on a Tuesday night, and nobody does that.

The information is public. grants.gov publishes every federal opportunity with its deadline, its award ceiling, and the list of who may apply. It is all there. It is just not in a form a person can ask.

Alexa+ speaks MCP. So I built the missing half: a self-hosted MCP server that turns that database into something you can ask out loud.

What it does

Six tools, each answering a question someone actually asks:

The repo also ships an Alexa+ simulator (web/) so the whole loop can be seen without an Echo on the desk: speak or type, watch it route to a tool, hear the answer — with the tool name, its arguments, the elapsed time, and the raw JSON on screen next to it. Everything in the demo video is a live call.

The part I actually care about: it refuses to guess

A wrong funding answer costs someone a weekend, or a filing fee, or a shot at rent money. So the honesty rules here are structural, not a line in a prompt:

• Source unreachable → returns error: source_unavailable and says "I have no numbers for you", never an estimate.
• A notice that publishes no applicant list (just Others (see text field…)) → reports status: see_notice. It does not say "you're ineligible". Turning a pointer-to-prose into a rejection would be the agent inventing a fact.
• An opportunity with no published deadline → counted and reported separately, never silently dropped from a "closing soon" list.
• The model that ranks fit → may set a score and a reason. It is never asked for a number. Every deadline and dollar figure is attached from grants.gov after the model has spoken, so a hallucinating model cannot invent money.

tests/test_radar.py::HonestyUnderFailure asserts each of those. They are tests, not promises.

How I built it

The transport, by hand. radar/server.py implements MCP Streamable HTTP (spec 2025-11-25) directly: one endpoint, initialize handing out a session id, an optional SSE stream, tools/list schemas generated from Python type hints and docstrings, tools/call returning both spoken text and structuredContent. No framework in the way — which made the spec easy to read off the wire while debugging, and means the repo has no dependency you have to trust.

The data, live and keyless. grants.gov search needs no API key. Responses cache to disk for six hours so a conversation stays fast and the source stays un-hammered.

The judgement, on your own machine. rank_for_me prefers a local model through Ollama. No key, no cloud account, no card — and the sentence describing your situation never leaves the house. That last part is not a footnote: the queries people bring to this are "grants for a family caring for a disabled child". Amazon Bedrock and plain keyword overlap are automatic fallbacks, and the engine that actually ran is reported in every response.

What I learned

Two date formats in one API. Search returns 10/09/2026; detail returns Oct 09, 2026 12:00:00 AM EDT. My first parser silently produced days_left: null for every detail lookup — the honest-looking failure is the dangerous one, because nothing errors. Now both are parsed and a test pins the long form.

"Not listed" is not "not eligible". My first eligibility check returned False whenever the applicant type wasn't in the published list. Then I read a real notice: the list said only Others (see text field entitled "Additional Information on Eligibility"). My tool would have told a qualified person to give up. That bug became the see_notice state.

Sequential model calls are a design flaw in a voice product. Judging eight opportunities one after another took most of a minute. Nobody stands in a kitchen for that. The fetches and the judging now run together.

Recording a demo is a privacy decision. My first screen capture grabbed the desktop — and swept in an unrelated window. I deleted it and rewrote the recorder to attach to Chrome over CDP and pull frames from the page itself. demo/capture.py is in the repo.

What's next

The weekly briefing is the shape this wants to grow into: not a thing you remember to ask, but something that has already looked while you slept, and tells you the one deadline that matters on Monday morning. weekly_briefing is the tool; a scheduled call is all that's missing.

Built With

• alexa
• grants.gov
• json-rpc
• llama
• mcp
• ollama
• python
• server-sent-events