为花生过敏朋友打造的 SafePlate：浏览器端过敏感知膳食规划器

My friend Maya has peanut, tree-nut, and sesame allergies. Every group dinner starts the same way: her squinting at

ingredient lists, googling "does pesto have nuts" (it does — pine nuts), and quietly eating plain rice while everyone else

shares satay.

So I built her SafePlate — an allergy-aware meal planner that reads every ingredient of every recipe for her:

• 🥜 Her profile, not a generic one — name, allergies, and how strict to be ("Sensitive" vs "Severe: flag every trace").
• 🍽️ A verdict on every recipe — All clear, Worth a closer look, or Not safe for Maya — with the exact ingredient that triggered it and why.
• 🕵️ It catches the hidden ones — pesto → pine nuts, Worcestershire sauce → anchovies, marzipan → almonds, tahini → sesame, green curry paste → shrimp paste.
• 🔁 Every flag comes with a swap — peanut sauce → sunflower seed butter, tahini → sunflower seed butter + lemon.
• 🏷️ "Check anything" scanner — paste any ingredient list (a sauce label, a snack bar) and it flags what's risky for her.
• 🗓️ One-click week planner — a full week of safe plates, saved locally on her device.

## Demo

👉 Live demo: https://karleeov-dev-hack.kimi.page

It's fully shareable by URL, which makes for a fun try-it-yourself:

• https://karleeov-dev-hack.kimi.page?friend=Maya&allergens=peanut,tree-nut,sesame — pre-loaded profile
• https://karleeov-dev-hack.kimi.page?scan=nutella%20spread,%20tahini%20drizzle,%20oat%20milk,%20eggplant,%20honey%20roasted %20peanuts,%20rice%20crackers — auto-runs the scanner on a tricky list. Three of those are dangerous for Maya; three only sound dangerous.

First visit downloads the model (~23 MB); after that it's cached and the whole app works offline.

## Code

https://github.com/karleeov/Maya

No build step, no framework — index.html, app.js, recipes.js, styles.css. The phrasebook, veto table, and swap

list in recipes.js are plain readable data, so adapting SafePlate for a different allergy set means editing a list, not a
model.

## How I Built It

The core is all-MiniLM-L6-v2, an open-weight sentence-embedding model (Apache-2.0), running entirely in the

browser

through transformers.js + ONNX Runtime Web. Ingredients and allergen profiles become 384-dimensional vectors;

cosine similarity is what knows marzipan lives suspiciously close to almonds.

But here's the honest part: embeddings alone were too noisy to trust with someone's health. I validated the scorer

offline against the exact quantized ONNX build the browser ships, and the false-positive ceiling was brutal — rice noodles
→ egg scored 0.80 (because "egg noodles" exist), tomato sauce → fish hit 0.76, butter → tree nuts 0.74. Pure semantic

search would have cried wolf about pancakes and stayed quiet about satay.

So SafePlate is a hybrid:

1. A curated, word-boundary-aware phrasebook per allergen carries precision — it knows eggplant ≠ egg and buttermilk ≠ butter.
2. The embedding layer rides shotgun — powering the "semantically close to X (matched term)" explanations, catching long-tail phrasings, and driving the caution tier.
3. A small veto table encodes domain truth the model can't know: coconut milk isn't dairy, oat milk is fine, peanut butter contains no butter.
4. Severity modes shift the thresholds, because "sensitive" and "anaphylactic" are not the same product. In Severe mode it still raises coconut as a tree-nut caution (the FDA lists it as one) — I'd rather explain a false alarm to Maya than apologize for a missed one.

The model downloads from the Hugging Face hub (with a mirror fallback), caches in the browser, and after that the app —

model included — runs with the Wi-Fi off.

t asking anyone's permission.

• Hackable. If her little cousin has a different allergy set, we edit a data list — not a billing plan.

The honest trade-off: a small open model is dumber than a frontier closed one. That's why the architecture is hybrid —

the open model does what it's good at (semantics, explanations, long-tail matching) and boring curated data does what it's

good at (precision). Open source didn't lower the ceiling; it changed where I put the walls.