Methodology
What 7 years of free-food posts say.
Before writing the first line of Swift, we parsed every MIT free-food post from December 2018 through May 2026. Here's what we found, what it cost, and how it shaped the product.
The dataset
4,489 individual free-food alerts, spanning December 2018 to May 2026
(7.5 years), pulled from MIT's Mailman archive at
mailman.mit.edu/pipermail/free-foods/. Each alert is one email forwarded
by a member of the MIT community announcing leftover food from a meeting, lab event,
seminar, or department gathering.
Note: the COVID period (March 2020 through summer 2021) is genuinely sparse in the archive — campus events stopped. Two posts in April 2020, two in May 2021. Any growth-rate analysis has to interpolate that gap.
Volume and growth
Posting volume has roughly doubled since 2019 — from ~50 posts/month then to ~100 posts/month in 2025-2026. Peak: 149 posts in November 2021, almost certainly the post-COVID return-to-campus catch-up. Current sustained pace is roughly 3-4 posts per weekday.
Pizza is not the dominant food
The single most surprising finding: pizza is only 5.8% of posts (260 of 4,489). Top 10 food categories combined account for just 36% of total volume; the remaining 64% spans hundreds of distinct cuisines.
| Category | Posts | % of total |
|---|---|---|
| (generic) "food" | 634 | 14.1% |
| pizza | 260 | 5.8% |
| Indian food | 147 | 3.3% |
| sandwiches | 130 | 2.9% |
| tacos | 91 | 2.0% |
| Thai food | 86 | 1.9% |
| Mediterranean | 61 | 1.4% |
| BBQ | 55 | 1.2% |
| Brazilian food | 52 | 1.2% |
| pasta | 45 | 1.0% |
The dietary distribution — the strongest argument for filtering
This is the finding that defines the product. The dot (●) marks tags appearing in fewer than 3% of posts — diets for which the unfiltered firehose is essentially noise.
| Tag | Posts | % of total | |
|---|---|---|---|
| vegetarian | 3,581 | 80% | |
| contains pork | 607 | 14% | |
| contains seafood | 368 | 8% | |
| vegan | 359 | 8% | |
| ● | gluten free | 129 | 2.9% |
| ● | halal | 82 | 1.8% |
| ● | lactose free | 57 | 1.3% |
| ● | contains nuts | 35 | 0.8% |
| ● | kosher | 34 | 0.8% |
| ● | nut free | 33 | 0.7% |
For any of these dotted tags — celiac, lactose-intolerant, nut-allergic, halal, kosher — the noise-to-signal ratio on the raw list runs between 33-to-1 and 142-to-1. The loudest example: kosher's 0.8% base rate at ~3-4 posts per day works out to roughly one kosher alert every 3-4 days, surrounded by ~99 irrelevant ones. That's why most observant students muted the list. Same math, slightly less extreme, for everyone else with a rare diet.
For the modal vegetarian, this app is a convenience — the firehose mostly works for them already. For the long tail of dotted diets — religious certifications, medical restrictions, life-threatening allergies — it's the difference between a list that's worth muting and one that's worth keeping on. That asymmetry is the entire argument for building this.
How the parser works
Each email is parsed by Claude Haiku 4.5 (Anthropic) using a structured-output "tool call" with this JSON schema:
{
"food_type": "pizza" | "Indian food" | ...,
"dietary_tags": ["vegetarian", "kosher", ...],
"building": "32" | "W20" | "14W" | ...,
"room": "G449" | "lobby" | ...,
"summary": "<120-char push-ready summary>"
}
The dietary policy is asymmetric by design: contains_* tags (allergens, pork, alcohol) are permissive — over-warning is safe. Positive claims (kosher, halal, gluten_free, vegan) are strict — only set when the email literally uses the word, because a false positive here can cause real harm.
What the parser does badly
Spot-checking 20 random predictions side-by-side with the original emails gave roughly ~70% strict-correct, ~25% acceptable-with-quibbles, ~5% genuinely wrong. The 5% wrong fell into a few recurring patterns:
- Sender's email-signature building bleeding into the food's location (now fixed)
- Letter-suffix building codes like "14W-111" being split wrong (now fixed)
- "Killian Court" vs "Lobby 10" conflation — both are at building 10 but spatially different (now fixed)
- Dietary over-claiming on positive tags — fixed by the asymmetric policy
A note on the data
The mailing-list archive itself is not redistributed — it's restricted to list subscribers, and we respect that. The aggregated counts and percentages on this page are derived statistics, not raw content.
Questions or want to bring this analysis to your university? Email hello@freefoodfinder.app.