网页 综合 更新时间 2026-09-07 06:57:48

The grant money already exists. My AI kept inventing foundations to spend it on - DEV Community

自动蜘蛛 dev.to 正文约 6000 字
The grant money already exists. My AI kept inventing foundations to spend it on - DEV Community
DEV Weekend Challenge: Generosity Edition Submission 💜
This is a submission for Weekend Challenge: Generosity Edition
The money exists. A small NGO just cannot find it.
That is what a generosity problem looks like at the small end. The giving has already happened — foundations with open, rolling, unclaimed programmes, sitting there — and it is spread across a few thousand pages nobody has time to read. After the 2025–26 collapse of USAID funding, organisations that had one funder now need six, and the people doing that searching are the same people running the programme: a director who is also the grant writer, working evenings. Generosity is not the scarce thing here. Attention is.
So an AI grant finder is an obvious idea. It is also a dangerous one, because the failure mode is not "unhelpful." A three-person NGO that spends a week writing an application against a deadline that never existed has lost a week it cannot get back, and it will not find out until it submits. The tool would have taken the one thing that was actually scarce.
FundFinderAI is the response, in one sentence: it searches the live web for currently open grants that fit your NGO, and then it refuses to trust its own model about any of them. Every application URL Gemini produces is independently fetched before you see it, and the card tells you what happened when we tried.
The interesting part is not that it searches. It is everything the app does to establish that the search actually happened and that the result actually exists.
Live: fundfinder-ai.vercel.app — describe an NGO, get grants, open a drafted Letter of Inquiry.
Give it 30–120 seconds. It is running ten to thirty real Google searches and then fetching every URL that comes back, and the page shows you the clock while it does. Paste this in if you would rather not invent an NGO:
NGO name: Kisumu STEM Girls Collective Location: Kisumu, Kenya Mission: We run after-school robotics and coding clubs for girls aged 12-17 in Kisumu, Kenya, and train their teachers to keep the clubs running. Focus areas: STEM education, girls education, teacher training Annual budget size: Under $250,000 annually Target population: Adolescent girls in low-income neighbourhoods of Kisumu
You will not get my results. They vary run to run — which is itself the point, and is in Limitations.
A real, unmodified run against the deployed app, for a girls' STEM education NGO in Kisumu, Kenya. Gemini ran 26 Google searches and returned five grants, sorted so the ones that survived checking lead:
The bottom row is the product working — and the grounding panel on that same page lists hansenfamilyfoundation.org among the pages it retrieved. Google's index has that funder. The live site does not answer. Grounding retrieved it and the URL is still dead — which is the entire argument for checking rather than trusting, in one row.
Ask for a Letter of Inquiry against a grant in that state and the app says so in red before you write a word: we could not reach this funder's page, so we cannot confirm this grant — or this funder — exists. Drafting a warm, professional letter to an organisation that may not exist is the failure this whole project is against, and the button that does it should not be silent about it.
An earlier run caught a harder one: au-eu-youthlab.com, a confident, plausible, entirely non-existent domain. DNS does not resolve it. It never reached the user.
When the link does check out, a second Gemini call drafts the letter — carrying its own standing caution, because a first draft is not a submission:
This is a post about plausible output not being proof, so the central claim — that the app searches rather than remembers — ships as a runnable check. It talks to the Gemini API and nothing else: not my server, not my code paths. About thirty seconds:
It prints which models actually honour googleSearch and exits non-zero if the one this app is configured with did not search. The table it produces is further down.
Live: https://fundfinder-ai.vercel.app
A grant radar for small NGOs. Describe your organisation once; Gemini searches the live web for currently open grants that fit, and drafts a tailored Letter of Inquiry for any of them.
Built for the DEV Weekend Challenge: Generosity Edition.
An AI grant finder that invents a grant is worse than no grant finder. A three-person NGO that spends a week preparing an application against a deadline that never existed has lost something it cannot get back.
So the model is not trusted about whether a grant is reachable. Every applicationUrl it produces is independently fetched before you ever see it and the result is shown on the card:
Four files carry the argument:
The premise is live Search grounding. The app was configured with gemini-3.1-pro-preview and passed tools: [{ googleSearch: {} }], and it returned six well-formed, plausible grants in sixty seconds. It looked like it worked.
It was not searching at all.
The response carried no groundingMetadata — the candidate had only content, finishReason, index. No error, no warning, no refusal. The model accepted a tool it does not run and answered from memory, in exactly the shape I asked for. The tell was in the output, not the API: that run invented an application URL.
This generalises further than one model. Sending one identical grounded prompt to four models:
Three of four accept googleSearch and never call it. Any of them would have produced a working, convincing, ungrounded app.
That check is committed as npm run check-grounding. It talks to the Gemini API and nothing else — not my server, not my code paths — and exits non-zero if the model the app is configured with did not actually search. If the premise of this project ever stops being true, that command fails.
My first version of that check asked whether groundingChunks came back non-empty. By that measure gemini-2.5-flash was also ungrounded, and I nearly changed the model a second time chasing it.
Look at the chunks column above: it is z
访问原文 返回搜索 https://dev.to/zedaina/the-grant-money-already-exists-my-ai-kept-inventing-foundations-to-spend-it-on-14po
本页为82搜索生成的网页快照,内容来自站内索引;演示数据的原文链接仅作示例。