FoodQCheck

FoodQCheck is my D4 Software Engineering capstone project at Politeknik Negeri Padang. It helps UMKM producers of dry snack products — keripik, kerupuk, biskuit, and kukis — run a quick visual quality screen from a smartphone photo before a product goes to market.

The system is deliberately scoped as a visual pre-screening tool, not a food-safety certification system. It does not replace chemical, microbiological, or laboratory testing. Instead, it answers a practical production question: does this product look visually sellable or not?

Key Features

1. Camera-to-result workflow: Capture or select a product photo, upload it, and receive a Layak Jual or Tidak Layak Jual decision with confidence score and short explanation.
2. Cloudinary-first image pipeline: The mobile client uploads photos to Cloudinary, then sends only the image URL to the inference API.
3. MobileNetV2 inference API: FastAPI downloads the image, runs preprocessing, classifies with MobileNetV2, and stores the result with model version and threshold metadata.
4. Detection history: Users can reopen previous results, review details, and manage their history.
5. Admin dashboard: Operators can inspect detection volume, review records, and inspect model-facing artifacts such as Grad-CAM heatmaps.
6. One React codebase, two clients: The same React app ships as a web SPA and an Android APK through Capacitor.

Technical Challenge

The hardest part was shipping a complete product loop under real capstone constraints: small visual dataset, server-side inference, mobile capture quality, and a result format that non-technical UMKM users can understand.

The solution was a clear end-to-end architecture:

React / Capacitor app
        ↓
Cloudinary image upload
        ↓
FastAPI receives image_url
        ↓
SSRF-guarded download + MobileNetV2 inference
        ↓
PostgreSQL stores detection history
        ↓
User result + admin review UI

On the model side, I used transfer learning with MobileNetV2, binary sigmoid output, and a production threshold of 0.3 chosen to prioritize recall for visually unsuitable products. On the product side, the API stores enough metadata — label, confidence, raw score, threshold used, model version, and explanation — so every result remains auditable after the demo ends.

Outcome

On the held-out test split of 72 images from a curated 509-image dataset:

• Accuracy: 91.67%
• Macro precision / recall: 91.74%
• F1-score: 91.67%
• Recall for Tidak Layak Jual: 94.29%

The software side covers authentication, detection, history, admin review, Docker deployment with Caddy and PostgreSQL, and API contract tests. The project has been presented as a complete capstone system.

Tech Stack

• Mobile / Web: React 19, TypeScript, Vite, Tailwind CSS, Capacitor Android
• Backend: FastAPI, SQLAlchemy, Alembic, PostgreSQL, JWT auth
• Machine Learning: TensorFlow, MobileNetV2, Grad-CAM, custom training/evaluation pipeline
• Infrastructure: Cloudinary, Docker Compose, Caddy, Gunicorn/Uvicorn
• Source:
  • Backend + ML: umkm-food-quality
  • Mobile client: umkm-food-quality-mobile