Research & Insights

Building ChatClothes taught me that multimodal AI lives or dies on orchestration. Dify coordinated an LLM agent, diffusion pipeline, and lightweight classifier across two free-tier GPU servers — turning three separate models into one conversational try-on system.

Six years running Spring Cloud across 5+ factory sites taught me: service boundaries that look clean on paper become debugging nightmares when a serial scale stops reporting and the production line stalls. Event-driven is powerful — but contract testing is not optional.

After 10 years of Android IM, I've touched every abstraction: WebSocket, Firebase, Easemob, native wrappers. For consistent sub-200ms delivery at 5K DAU, a custom NDK TCP/UDP transport layer remains the only reliable answer.

Hundreds of factory workers using Android terminals in noisy, gloved, time-pressured environments rewired how I design mobile interfaces. Large touch targets, offline-first, serial bridge for weight auto-fill — these aren't preferences, they're survival.

Training YOLO12n-LC for garment classification: 5MB model (8× smaller than the 40MB base), 94.2% accuracy, 35ms inference on Raspberry Pi 5. Channel reduction + ARM quantization — not architecture search — got us there.

Leading a 6-person cross-functional team across time zones taught me: clear ownership, shared demos every week, and protecting deep work time matter more than any methodology. Async-first coordination is a feature, not a compromise.