NTU Master of Computing in Applied Artificial Intelligence student working on trustworthy retrieval-augmented generation, especially citation faithfulness and evidence repair for multi-hop question answering.
My current work is organized around one question: how can AI systems make their evidence, reasoning, and failure modes easier to inspect and repair?
- Citation faithfulness and evidence repair in fixed-answer multi-hop RAG
- Lightweight verifier and reranker methods for evidence-grounded generation
- Explainable AI for model diagnostics and failure analysis
- Reproducible applied AI workflows for research and competitions
- l40s-llm-bench - primary open-source maintainer project: reproducible LLM-serving benchmark scaffold with
v0.1.4, passing CI, reviewer smoke proof, Codespaces-ready maintenance path, result-submission example bundle, result-review checklist, and a dry-validatable vLLM/L40S smoke profile. - FaithfulRepair - working manuscript on evidence-insufficient citations in multi-hop RAG. A public-safe repository shell has synthetic repair, verifier, distractor, and trace experiments.
- BirdCLEF+ 2026 XAI - active bioacoustic AI competition workflow with synthetic submission, explanation-report, and release-checklist utilities.
- ARC XAI Reasoning - exploratory explainable reasoning baseline with synthetic transformation logs, failure taxonomy, rule coverage, and explanation audits.
Other project links will be added after the relevant repositories are public, verified, and safe to release.
| Project | Current signal | Next action |
|---|---|---|
| l40s-llm-bench | Public flagship; v0.1.4 release, CI, reviewer smoke proof, Codespaces path, community submission/review path, vLLM/L40S smoke profile |
Collect real tester feedback or first hardware-backed smoke artifact |
| FaithfulRepair | Strong research direction, paper-sensitive | Keep public package synthetic until manuscript strategy is stable |
| ARC XAI Reasoning | Useful explanation sandbox | Add public-safe rule composition toy cases |
| BirdCLEF+ 2026 XAI | Competition-supporting utilities | Keep real notebooks private and publish only synthetic diagnostics |
I keep a local GitHub portfolio signal model to decide which projects are ready to publish or pin. It scores reproducibility, evidence artifacts, community utility, technical depth, focus alignment, release safety, and profile clarity. It also ranks the next small optimization experiments by comparing my projects against patterns borrowed from strong AI and open-source GitHub profiles.
See docs/github_profile_inspiration.md for the inspiration scan and modeling
logic.
Before moving into AI, I worked in finance, audit, and risk advisory. That background shaped my interest in traceability, validation, and reliable decision support.