LLM evaluation · agent reliability · post-training research engineering
Hong Kong · The University of Hong Kong · Mathematics (second major)
Résumé · Research & engineering portfolio · LinkedIn · ORCID
I am an HKU undergraduate studying Economics & Finance with a second major in Mathematics, and a research engineer working on reproducible ways to evaluate LLM behavior, preserve executable state in agent systems, and test post-training claims. I currently contribute to AI agent development and evaluation at Tencent and conduct research at HKU's Centre of AI, Management and Organization.
- Research: two sole-authored papers accepted to non-archival workshops at COLM 2026, on post-training trace reliance and executable agent state.
- Open source: 11 merged pull requests across nine AI/software repositories as of August 2026, including OpenAI Agents, OpenAI Node, Pydantic AI, OpenCUA, LightRAG, RAG-Anything, Agno, txtai, and xberg.
- Research engineering: public, runnable evaluation artifacts with smoke checks, result cards, counterfactual controls, and explicit claim boundaries.
Accepted poster, Context Beyond the Window — COLM 2026 Workshop (sole author; non-archival).
Tests whether agent handoffs preserve the bindings required to resume work—not merely identifiers or visible text—and packages the failure cases as an executable reliability gate.
Accepted, AIMS — COLM 2026 Workshop (sole author; non-archival).
Uses interventions and controls to distinguish behavioral reliance on supplied traces from surface-level imitation of trace form.
My contributions focus on evaluator correctness, tool-call identity, multimodal data handling, retrieval checks, schema behavior, and deterministic tests.
Selected merged work:
- openai/openai-agents-python#4361 — honor WAV sample width when decoding audio.
- pydantic/pydantic-ai#5355 — preserve xAI tool-result IDs.
- xlang-ai/OpenCUA#56 — fix AgentNetBench evaluator scoring edge cases.
- HKUDS/LightRAG#3038 — add an offline sample-retrieval check for RAGAS evaluation.
- HKUDS/RAG-Anything#273 and #278 — preserve multimodal aliases and correct duplicate detection.
- openai/openai-node#1903 — sanitize Zod definition references.
All 11 merged upstream pull requests
- openai/openai-agents-python#4361 — audio sample-width handling
- xberg-io/xberg#1413 — slide page-number preservation
- neuml/txtai#1183 — OpenAI chat-model selection
- openai/openai-node#1903 — Zod definition-reference sanitization
- agno-agi/agno#8145 — RemoteTeam knowledge-filter attributes
- pydantic/pydantic-ai#5533 — deterministic
Retry-Aftertesting - HKUDS/RAG-Anything#278 — duplicate detection
- pydantic/pydantic-ai#5355 — xAI tool-result identity
- HKUDS/RAG-Anything#273 — multimodal content-list aliases
- HKUDS/LightRAG#3038 — offline retrieval evaluation
- xlang-ai/OpenCUA#56 — evaluator scoring boundaries
- Evaluation harnesses for LLM and agent behavior
- Reliability checks for tool use, state transfer, and data pipelines
- Post-training diagnostics with intervention-based controls
- Reproducible Python research infrastructure and regression tests
The standard I aim for is simple: a concrete failure mode, a minimal reproduction, a meaningful control, a runnable check, and a precise statement of what the evidence does—and does not—establish.
The best public way to reach me is LinkedIn. I do not publish personal email or phone details on this profile.