A charity does not close because its overhead was too high. I checked 328,186 tax returns.

This is a submission for Weekend Challenge: Generosity Edition

I have given money to charities for years and I have never once known whether the one I gave to was about to close. The number everybody puts in front of you is overhead, the share of spending that does not go to programmes. Every rating site leads with it. I assumed it meant something.

So this weekend I downloaded the IRS Statistics of Income extracts of Form 990 and Form 990-EZ, which is 3,697,515 filings covering 700,873 organisations across fiscal 2017 to 2024, then asked a narrower question. Not whether a charity is efficient. Whether it is going to still be there.

I took the 328,186 organisations that filed a long-form Form 990 for fiscal 2018 or 2019, then checked which of them appear on any filing for fiscal 2022 or later. 8.88% do not. Then I scored two candidate signals inside each of ten spending deciles, so that nothing I found could turn out to be a story about big charities outliving small ones.

Feeding America spends $4.93 billion a year and holds 1.04 months of it in cash. The Greater Chicago Food Depository spends $261 million and holds half a month. Second Harvest Food Bank of Central Florida holds nine days. None of that is on any rating site, because the number nobody computes for you is the one that decides whether the lights stay on.

What I Built

Keepalive computes months of runway for 579,178 US nonprofits, straight off their own filings. Runway is cash plus savings divided by one month of spending. It is the number a finance director lives by and the number a donor never sees.

Across 1,573,687 long-form filings the median is 5.5 months. 33.5% hold under three months. 15.1% hold under one.

The site does four things with that.

1. Where does my gift buy the most time. Say what you can give, pick a cause and a state. It ranks real organisations by days of runway your money adds, which is 365 * yearly gift / annual spending. $25 a month is 12 hours of runway at a $260 million food bank and 9.4 days at a $23,000 one. The same money, three orders of magnitude apart in effect. That single line changed how I give.
2. Look up any of the 579,178 by name or EIN, with seven years of revenue, spending and cash on one chart.
3. Read the brief. Gemini writes it from a fixed packet of that filer's own figures and is handed nothing else, so there is nothing for it to invent. It has to return a line naming what a tax filing cannot tell you.
4. Check it. One click publishes the SHA-256 of the figures on screen and the brief just written as a memo on Solana devnet, so what Keepalive said about an organisation on a given date lives somewhere its author cannot quietly edit.

The finding

Both tables below hold inside every one of ten spending deciles.

Runway predicts survival. Thin runway is the worst of the three bands in all ten deciles. Pooled with deciles weighted equally: 86.79% of the under-three-months group was still filing, against 92.61% at three to twelve months and 93.56% above a year. The gap is 10.3 points in the smallest decile and 1.6 in the largest, so it bites hardest exactly where the money is smallest.

Fundraising thrift points the wrong way. The organisations that spent under five cents to raise a dollar survived less than the five-to-fifteen-cent band, in all ten deciles: 91.16% against 94.86%. Spending more on fundraising went with surviving more. In nine of the ten deciles the thriftiest band is the worst of the three outright. Only one of these two numbers is on the rating sites. It is the one that is backwards.

That is the starvation cycle, visible in tax returns. Judge a charity on how little it spends on itself and you reward the ones with no capacity to raise next year's money.

Demo

Live: https://keepalive.zkasuran.dev No sign-in, nothing to install.

Try it in 30 seconds:

1. Open the site. The two survival tables under The finding are the whole result, drawn from the same query.
2. In the search box type greater chicago food depository. Open it. The runway card reads 0.5 months on a $261m budget. The peer card shows 95.6% of organisations in its own spending decile and runway band still filing.
3. Scroll to The honest brief and wait about eight seconds. Gemini writes it live. Read the last line, the one that starts "What these filings cannot tell you".
4. Press Publish a checkable receipt. About four seconds later you get a Solana devnet signature. Open it. The memo on chain holds the SHA-256 of exactly what you just read.
5. Back at the top, set the amount to 25 per month, cause Food, agriculture and nutrition, state TN. The list re-ranks by days of runway your $300 a year adds.

Code

zkasuran / keepalive

Months of runway for 579,178 US nonprofits, read off their own IRS filings

Keepalive

A charity does not close because its overhead was too high. It closes because it ran out of months.

Keepalive reads 3,697,515 IRS filings and computes, for 579,178 US nonprofits, the one number no rating site puts in front of you: how many months of spending they hold in cash Then it tells you what your gift does to that number.

Live: https://keepalive.zkasuran.dev

Built for the DEV Weekend Challenge: Generosity Edition 4 to 7 September 2026. Everything in this repository was written inside that window.

The finding

Two candidate signals, measured on the same 328,186 organisations that filed a long-form Form 990 for fiscal 2018 or 2019, scored on whether the same EIN appears on any filing for fiscal 2022 or later. 8.88% do not.

Months of runway predicts survival. Under three months is the worst of the three bands in all ten spending deciles, with no exception.

…

View on GitHub

npm run data reproduces every number: it pulls the IRS extracts, loads 3.7 million filings, prints the finding and writes the static shards the site reads. It needs no credentials and no account.

How I Built It

There is no runtime database. The site is 4,542 static files on a CDN plus four small functions that hold keys a browser cannot have.

Search is a static inverted index. 579,178 names tokenise to 1,910,797 postings, hashed into 2,048 shard files. The browser hashes your query tokens, fetches only the shards it needs and intersects them. Typing "greater chicago food depository" costs three fetches and no server. Each posting list is ordered by spending, which matters more than it sounds: truncate a common token like "food" and the intersection silently stops finding real names. That was my first real bug.

Organisation records are sharded by EIN modulo 1024, so one lookup is one 250 KB fetch that the CDN then holds.

Google AI writes the brief under a JSON schema, from a packet containing every number it is permitted to mention and no others. Absent fields are dropped rather than sent as zero, because "not reported" and "$0" are different claims and a model handed a zero will tell you the charity has no employees. The system prompt forbids computing a new ratio. One measured detail: Gemini 3.5 Flash spends its thinking tokens out of the same maxOutputTokens budget, about 2,200 of them here, so a tight cap truncates the JSON and the parse fails. That looks exactly like a bad model and is not.

Solana is the only place the receipts live. Keepalive keeps no copy. Remove Solana and the app can no longer prove that today's number is the number it computed from the filing, which is the honest answer to "how do I know your AI did not make this up".

The transaction is built against the wire format rather than with a client library. A memo transaction is one signature, two account keys and one instruction. Ed25519 is already in crypto.subtle, so an SDK would have added megabytes to buy nothing. test/attest.wire.test.mjs proves those bytes equal the ones @solana/web3.js produces: message, signature and full serialized transaction, byte for byte, including a memo that crosses the compact-u16 boundary.

One thing I want to state plainly rather than hide. The signature is made on the server, where the key is. The transaction is submitted by your browser. That is not a design flourish. The public devnet RPC answers Cloudflare's egress with 403 Your IP or provider is blocked from this endpoint while a visitor's own connection is fine. That RPC allows any origin. So the key stays server side and the network call happens where the network works. The only thing a client controls is the blockhash, which is length checked and which cannot turn a memo into a transfer.

Snowflake is the warehouse path, in data/snowflake.sql: the same cohort plus one statement that cuts it into deciles then scores both signals inside every decile at once. Cortex does the language work next to the data, with AI_CLASSIFY over names the IRS only gave a letter code and AI_AGG writing one sentence per filer from all of its returns together. There is a licensing line I held to: Marketplace terms do not permit redistributing a dataset, so every per-organisation figure this site serves comes from my own public domain IRS download instead. Only aggregates ever leave a warehouse. Sources and the exact granting sentence for each are in DATA-SOURCES.md.

ElevenLabs reads it out, because the reason a lot of people cannot use a giving tool is that it is a wall of financial text. The free tier is about ten minutes of audio a month, so architecture had to do the work rather than budget: everything that never changes is pre-rendered once at build time by tools/narrate.mjs into static mp3. A visitor who plays the tour or the finding spends nothing. Only a brief about one specific filer is synthesised live, cached in the Worker by a SHA-256 of the text, so the second person to read the same organisation is free too. You can watch it happen: the response carries x-keepalive-audio: fresh the first time and cached after.

The tour is in English, Hindi, Tamil and Spanish. Gemini does the translation with an instruction to keep every figure exactly as stated. ElevenLabs speaks it. The English stays the source of truth. That is the one part of this build I would call generous rather than clever. A US tax filing is hard to read in your first language. The people most likely to be deciding about a small charity back home are not all reading English.

Two things that will bite anyone doing this on a free key. A scoped key answers 401 missing_permissions on GET /v2/voices, so a voice id has to be configured rather than discovered, with a short list of defaults to fall back through. Free accounts also get 402 paid_plan_required on library voices, so a voice id copied out of a blog post is a dead end. The four in the list are ones I verified synthesise on this key.

What is real and what is not

• Every figure on an organisation page is that organisation's own reported number, labelled with the fiscal year, computed in your browser from a static file. Nothing is modelled and nothing is imputed.
• Form 990-EZ does not break out cash, so EZ filers show no runway rather than a guess.
• The survival tables are associations, not causation, controlled for size by decile and for nothing else.
• Solana devnet, not mainnet. No money moves anywhere in this project. Keepalive is not a payment processor and never touches a donation.
• Keepalive is not a charity rating. It can tell you what an organisation's finances look like. It cannot tell you whether they are good at the work. Neither can any rating site.
• AI assistance (Claude) was used in building this. The design, the analysis and the verification are mine. Every number above came out of SQL I can rerun in front of you.

Prize Categories

• Best Use of Google AI. Gemini writes every brief under a JSON schema from a packet of that filer's own figures, with a rule that a missing field means not reported. It also writes the narration translations without touching a single figure.
• Best Use of Solana. A devnet memo is the only store of record for what this site published. Remove it and no claim on the page is checkable. The transaction is hand built and proved byte-identical to the reference implementation.
• Best Use of ElevenLabs. The tour and the finding are pre-rendered at build time in four languages, so listening costs no credits and needs no key. A brief about one organisation is spoken live, then cached by a hash of the text.

The Snowflake path is written and in the repository as data/snowflake.sql, and the live deploy tells you plainly that its credential is not bound: /api/status returns snowflake: false. The finding tables you are reading were produced by the same SQL run locally in DuckDB rather than in a warehouse. I would rather say that than claim a category a judge cannot see working. If it changes before the deadline I will say so here in an edit, with the timestamp.

Thanks for a theme worth building for. The most useful thing I learned is that the kindest gift to a small charity is not the largest one, it is the predictable one. The number to ask about is months.