DecayDock Uses Local AI to Track Fridge Freshness

Hackster.io publishes coverage of an open DIY project called DecayDock, an AI-powered fridge companion built by Instructables user ptallthings93. Per Hackster, the design centers on an ESP32-CAM development board with a built-in camera and Wi-Fi, a front-mounted TFT display, and a custom magnetic enclosure that attaches to a refrigerator door. Hackster reports that the project runs its trained ML model locally using Edge Impulse rather than relying on cloud inference. The system detects common grocery items, Hackster lists tomatoes, onions, bananas, spinach, milk cartons, and leftovers, adds them to a digital inventory, and shows freshness as color-coded progress bars (green, yellow, red). Hackster also reports that the author trained the model using dozens of photos captured in real kitchen lighting with the same ESP32-CAM hardware.

Editorial analysis - technical context: Projects like DecayDock typically combine tiny convolutional classifiers or quantized mobile models with on-device preprocessing to fit within the memory and compute limits of microcontrollers such as the ESP32-CAM. Using Edge Impulse for dataset collection, training, and model optimization is a common workflow because it automates quantization and converts models to microcontroller-friendly formats. Capturing training images in the intended deployment environment, as the creator did, reduces domain shift and improves real-world detection accuracy on low-capacity hardware.

This build is part of a broader wave of practical edge-AI hobby and maker projects that push useful capabilities onto low-cost devices. For practitioners, DecayDock is a compact example of embedding vision, simple inventory tracking, and heuristic freshness estimation without recurring cloud costs or privacy tradeoffs tied to offloading images to external servers.

Editorial analysis: Observers may look for reproducibility details on the Instructables page (model architecture, input resolution, quantization settings, and inference latency) and for community forks that add barcode scanning, thermal sensors, or integration with home automation platforms. Per Hackster, the original author documented wiring and enclosure design, which lowers friction for replication by other makers.