M.Sc. in AI and Entrepreneurship, HKUST · B.Eng. in Computer Science and Technology, ECUST
I build LLM and agent systems, and the evaluation harnesses that keep them honest. My work sits between applied AI engineering and measurement: retrieval-augmented pipelines, structured outputs and tool calling, and benchmarks that test whether a model actually understands a situation rather than pattern-matching it.
A counterfactual image-understanding benchmark for vision-language models. It separates "the model recognises the picture" from "the model can reason about what would happen if one variable changed", using paired relevant and irrelevant knowledge conditions to isolate where accuracy gains actually come from.
Natural-language maintenance ticket triage: unstructured repair reports become categorised, geocoded, worker-assigned work orders. A pure-LLM pipeline was not reliable enough, so the LLM is constrained by a rule engine at every step. Dockerised FastAPI backend, Vue front end, public demo.