At Lenovo, I built SelfTeacher, a self-distilled post-training pipeline that improved Qwen3-4B from 58% to 82% on AMC23 and from 73% to 87% on MATH-500 over the SFT baseline. The training loop runs the 4B student without a runtime teacher, cutting GPU deployment cost by 10x.
I implemented an end-to-end GRPO stack for RL post-training using PyTorch, TRL, PEFT, and GPU-colocated vLLM sampling. I also co-invented a sentence-level log-probability difference reward, granted as a CN national invention patent.
At the Ministry of Public & Business Service Delivery and Procurement, I refactored legacy workflow logic into a modular Java and Spring Boot architecture. I also improved Selenium automation resilience and integrated DB2 records for test data processing.
At Intellectual Property Association, I analyzed filing data from nearly 2,000 local enterprises and built a Java and Spring Boot repository of 2,000+ IP records. My projects include a student management platform adopted by the Business School for remote education and a digital human assistant with speech recognition and response generation.
