At MetaGPT (DP Technology), I analyzed execution traces across 100+ multi-agent sessions to identify failures in orchestration, tool use, and memory retention. Those findings informed prompt redesign and template improvements, and I designed 30+ end-to-end showcase prompts adopted as flagship product demos.
At Fyde Innovations (FydeOS), I built an end-to-end LLM inference pipeline for a tablet using the Rockchip RKLLM toolkit. I also developed Python FastAPI serving infrastructure with real-time token streaming and reduced cold-start latency by 55% with a CPU pre-ramp synchronization script.
