FreshVision
Pre-Visual Food Waste Prevention
Don't detect waste. Prevent it before it happens.
— notebook walkthrough of the risk engine, image heuristics and simulator
Live: https://freshvision-nine.vercel.app/ · Demo video: https://youtu.be/4cgehACaUok
FreshVision is a computer vision and inventory intelligence platform that identifies potentially at-risk food batches and helps users simulate waste-prevention strategies.
Built for Problem Statement 4 — Predicting the Unseen: Pre-Visual Food Waste Mitigation.
Features
- Inventory dashboard — produce batches ranked by risk, with quantity, storage age, conditions, consumption rate and status (demo data, clearly labelled).
- 3D inventory twin — interactive three.js storeroom (React Three Fiber) with low-poly walk-in chiller, reach-in fridges, produce cooler and ambient rack; produce crates on pallets coloured by projected waste, safety holds, temperature sensors and FreshVision camera fields of view. Toggle baseline vs intervention scenario; click a pallet to inspect and simulate.
- Image inspection — upload a produce photo; colour-distribution, discoloration and texture features are extracted locally in the browser and shown as an annotated overlay. Optional MobileNet (TensorFlow.js) identifies the produce type only.
- Explainable risk engine — transparent weighted scoring of storage age, temperature and humidity deviation, demand shortfall and visual indicators, with per-factor contributions and configurable weights. The visual score is kept separate from the overall inventory-risk score.
- Counterfactual simulator — compare a baseline plan with interventions (menu prioritisation, storage correction, redistribution) and see the change in projected quantity at risk.
- Actionable recommendations — prioritised, rule-based actions. Food-safety holds always override scores; held stock is never recommended for use.
All scores are heuristic estimates and all scenarios are model outputs — not validated forecasts, measured waste avoided, or food-safety determinations.
Run locally
npm install
npm run dev
Open http://localhost:5180
Production build
npm run build
npx vite preview --port 4180
Deploy (Vercel)
Static frontend, no backend or API keys. vercel.json sets framework vite, install npm ci, build npm run build, output dist.
npx vercel
Tech
React 18 · Vite 5 · TypeScript · three.js + React Three Fiber 8 + drei 9 · Canvas 2D image processing · TensorFlow.js MobileNet (lazy-loaded, optional)
Limitations
- Risk weights, Q10 ageing factor and shelf lives are uncalibrated assumptions.
- Image heuristics can be affected by shadows, glare, packaging and naturally multi-coloured produce.
- MobileNet weights download from tfhub.dev on first use (internet required).
- State is in-memory; reloading resets to the demo inventory.
