# Minh Trinh

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Minh Trinh](https://cerebralvalley.ai/u/-minhtos-)
- **GitHub:** https://github.com/minhtos/meal-measure
- **Demo video:** https://www.loom.com/share/c56ad021969f48bc95c4ec595afff54f
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3427

meal-measure is a photo food-measurement and logging tool. SAM 3.1 segments the plate, Muse Spark identifies each food and estimates its portion as a range, and plain code — never the model — computes calories and macros from USDA and Open Food Facts entries.

Anyone who has logged meals knows the friction: search a database for every item, guess every portion, repeat for every meal. It takes minutes of homework per meal, so most people quit within weeks. I know this problem firsthand, and I wanted the measurement without the homework — something that frees up time and mental space instead of consuming both.

meal-measure is a measurement instrument, not a nutrition coach. It has no targets, no goals, no advice, no "healthy/unhealthy" judgments, and no diet plans. It does one thing: turn a photo of a meal into a checked, sourced measurement — foods identified, portions estimated as ranges, nutrients computed from database entries — that you can correct in seconds and log.

What separates it from a typical AI nutrition app:
1. SAM 3.1 segmentation first. The plate is cut into items with real mask outlines the user can see and tap — not a single black-box guess at a whole photo.
2. Scale-anchored portions. One optional whole-plate weight, read off the user's kitchen scale by OCR and confirmed by the user, anchors the estimate and visibly narrows every range. The vision model is explicitly blind to scale displays; the reading enters only through the confirmed OCR step.
3. USDA-computed values, shown as ranges. The model names foods and estimates grams. It never outputs a calorie or macro number — code multiplies confirmed grams by database nutrients per 100 g, and every value carries its source entry.
4. Published measured error. I weighed seven real meals component-by-component and published the audit, misses included: median total error ≈ 31%, 4 of 7 totals inside the estimated range. Known failure modes are named, and the correction UI exists precisely because the audit showed where it fails.

How it works: photo → SAM 3.1 segments items and draws mask outlines → Muse Spark identifies each item, estimates a gram range, and picks a database match (top-5 near-miss candidates attached to every item) → optional anchors: user-confirmed scale OCR, brand/place context signals, or a barcode lookup against a local Open Food Facts index (814k US products) → code computes nutrients → one-tap corrections recompute instantly. Every runtime model is a Meta model.

All application code was written during the hackathon (Oct 26 – Nov 4). Pre-existing pieces, declared in the repo's PROVENANCE.md: the public datasets (USDA FoodData Central; Open Food Facts) and a generic project scaffold.

## More from Meta Global AI Developer Hackathon

- [LosCommosHermanos](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3424)
- [LinguaCheck](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3425)
- [Kleos9](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3426)
- [OwlOwlOwl](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3428)
- [LUMI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3429)
- [Minsk](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3430)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3427. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt