🎓 Third-year Ph.D. Candidate at City University of Hong Kong
💻 Code Generation & Evaluation · 🛠 Software Testing & Reliability · 📈 Time Series & Information Theory
Building executable systems with inspectable and reproducible results.
- Code generation & evaluation · Benchmarking LLMs and coding agents on SWE-bench tasks and quantitative trading code, with execution-based evaluation and iterative refinement.
- Software testing & reliability · Failure analysis, test-scenario amplification, and regression evaluation.
- Time series & information theory
| Project | Focus | What's inside |
|---|---|---|
| RouteBR | Constrained LLM routing | Deterministic pre-response controller, a 600-scenario benchmark, and reproducibility records. |
| Bug2Scenario | Autonomous-driving software testing | Root-cause-guided scenario amplification, simulator integration, and execution-based preservation assessment. |
| SpectralCSTE | Transfer entropy & time series | Gaussian CS-TE experiments, numerical audits, frozen configurations, and reproducible outputs. |
| HGARena | Repository-level coding agents | Graph-assisted retrieval and submitter–reviewer–executor search; a public case with success and failure records. |
| SysTB | Quantitative trading code & evaluation | Strategy-to-code generation, sandboxed backtesting, validity checks, cross-model evaluation, and iterative repair. Currently under revision. |
Turning research ideas into systems that can be inspected, tested, and reproduced. I value explicit assumptions, useful failure cases, and evidence that makes the limits of a method clear.