Research & Insights

After deploying ChatClothes on Raspberry Pi 5, I'm convinced the future is local-first: privacy, latency, and cost all favor edge inference. Cloud is for training; edge is for serving.

Six years running Spring Cloud across 10+ factories taught me: service boundaries that look clean on paper become debugging nightmares on the factory floor. Event-driven is powerful but demands rigorous contract testing.

After 10 years of Android IM, I've tried every abstraction: WebSocket, Firebase, native socket wrappers. For true sub-200ms delivery at scale with 5K DAU, a custom NDK TCP/UDP transport layer remains unbeatable.

Hundreds of workers using shop-floor apps in noisy, gloved, time-pressured environments changed how I think about mobile design. Large touch targets, offline-first, scan-centric workflows — not optional.

Compressing YOLO12n to YOLO12n-LC gave 94.2% accuracy at 8x smaller size — but the real win was <10s inference on Raspberry Pi 5. The lesson: quantization + distillation > architecture search for edge deployment.

Managing a distributed team across China and New Zealand taught me: explicit contracts, shared rituals, and protecting focus time matter more than any methodology. Async-first communication isn't a constraint — it's a feature.