# Broken By Design

- **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:** [Svetlozar Argirov](https://cerebralvalley.ai/u/zaro), [Bhumika Sharma](https://cerebralvalley.ai/u/schmithvillers)
- **GitHub:** https://github.com/Br0kenByDes1gn/hold-that-cart
- **Demo video:** https://drive.google.com/file/d/1MPK3Z0w9-RXxzeabH2rMEjYbRbVTfS1h/view
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2000

Hold That Cart is a PWA that turns receipts into a derived household pantry and intervenes at the moment of purchase. Its core is a vision model that reads photographs of receipts, shopping carts, and pantry shelves, extracting items into a strict schema rather than returning free text: name, brand, pack size, quantity, price, and a fixed category.

A second model call supplies the product knowledge the photo cannot contain — what the item is, how long it keeps, whether a better form exists (a refill, a larger pack, a store brand, wholegrain, a reusable instead of a disposable), and a directional estimate for money, health, and planet. That enrichment runs once per product and is cached across households, so it stays cheap and consistent; shoppers are never sent a fresh model call mid-aisle.

The AI is also the input layer. Dictated grocery lists are parsed by the model into named products with spoken brands stripped into a separate field, and camera detections are grouped into single lines with summed quantities. Text input is handled locally and never calls the model.

Two constraints shape the design. Grounding: the enrichment schema has no price fields and model output containing prices is rejected, so analysis can inform but never claim a saving. Confidence: health and planet signals are shown only above a threshold, and an indicator without support stays grey.
What the model produces is context, not decisions. It describes a product and estimates its properties; a deterministic engine compares that against the household's own purchase history to produce the buy, skip, and swap view. The split is deliberate — language models handle interpretation, which varies and cannot be audited, while anything a user acts on is reproducible code.

Built with TypeScript throughout — React and TanStack on the client, Hono and Drizzle on the server, Muse via the Meta Model API, Postgres dialect in both development and production. Accounts are optional; a device-scoped household exists from first use.

## More from Meta Global AI Developer Hackathon

- [famuse](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1997)
- [Meta](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1998)
- [Karthi_Sk_Sakthi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1999)
- [Muse Shopping](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2001)
- [WeaveWithAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2002)
- [Someday](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2003)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2000. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt