Computing Life · YageAug 2, 2026, 14:00 UTC+8

10-hour sprint: AI ships a working neural network on an ESP32-CAM to detect garage door state

The author paired an ESP32-CAM with GPT-5.6 SOL to build a garage-door sensor in 10 hours. The AI self-labeled asymmetric data, fine-tuned MobileNet V4, and switched to LSQ-based QAT after int8 quantization caused >30% classification flips—final AP drop was only 8 points. Inference takes 200 ms, image capture 400 ms; the device wakes every 10 minutes and sends results over WiFi. The author's main contributions were enabling a closed-loop dev cycle for the AI and flagging data imbalance and quantization pitfalls upfront. Total cost is not disclosed.

Why it matters: A solid embedded AI hands-on piece with real model choice, quantization failure, and latency numbers — not a generic tutorial. Downside: niche topic (personal DIY) and the article body cuts off mid-quantization detail. Featured because all three HKR axes hit, but importance st...

Read the original ↗Export Markdown

Take

This one's worth a click because it pushes past "AI writes code" into "AI makes engineering calls." The author gave high-level direction—data imbalance, quantization will hurt—and GPT-5.6 SOL handled the rest: self-labeling with a local Vision LLM, picking MobileNet V4, spotting the >30% classification flip after naive int8 quantization, then switching to LSQ-based QAT to recover most of the accuracy. Inference runs at 200ms, image capture takes 400ms, and the device wakes every 10 minutes to send results over WiFi.

I'd discount the "10-hour" headline a bit. This is a heavily overfit single-scene task—fixed camera position, limited lighting variation—so it doesn't generalize to production products. The useful bit is the closed-loop dev setup: the author deliberately had the AI build a test firmware that self-flashes, reads serial output, and iterates without a human in the loop. That's normally wasted effort in traditional dev; here it's the leverage point.

Total cost and long-term stability aren't disclosed. The power-bank cutoff warning is a nice real-world detail. If you're doing on-device deployment, the QAT selection logic and the closed-loop design are what to steal—not the speed number.

Source: Computing Life · Yage · yage.ai
Published: Aug 2, 2026, 14:00 UTC+8 · Aug 2, 2026