cv

Research Engineer

SMU SOAR Lab

Research on LLMs for software engineering under Prof. David Lo.

• First-author research on inference-time control of coding agents and code-model security (ASE 2025; EASE 2026 → IEEE Software; one paper under review at a CORE A* venue).
• Own projects end-to-end: empirical study, dataset construction, model training & fine-tuning, evaluation, and research tooling.
• In collaboration with GovTech Singapore (from 2026): repository-level agentic vulnerability detection — code property graphs plus an LLM-as-a-judge context-extraction stage.

Ph.D. (part-time)

Singapore Management University

Computer Science — Software Engineering & Trustworthy AI

• Advisor: Prof. David Lo (SOAR Lab)

B.Eng.

Singapore University of Technology & Design

Computer Science & Design

• Minor in Artificial Intelligence
• Exchange: NUS School of Computing (2021)

SCIS Research Excellence Award — Tier 1

SMU School of Computing and Information Systems

Awarded for a first-author A* publication (50% tuition-fee waiver).

Best Student Award (Term 3) — ranked 1st of 400+ students

Singapore University of Technology & Design

Top-ranked student in the cohort.

Honours with Highest Distinction

Singapore University of Technology & Design

GPA 4.83/5.0; SUTD Honours List (multiple terms).

Senior Middle 2 Scholarship

Singapore Ministry of Education

Selected for NUS Exchange (SUSEP)

NUS School of Computing

One of 7 students selected from the pillar.

Fail-Fast, Restart-Smart: Early Failure Prediction and Restart for SWE Agentic Tasks

Under review, 2027 · CORE A*

A two-stage controller with two complementary modes. Fail-Fast (cost): a 0.6B monitor early-terminates doomed runs — 14.6–20.4% token savings at a 5% false-positive rate, transferring across four policies including a closed-API model. Restart-Smart (quality): on an alarm, relaunches a fresh rollout with the prior repo diff as an optional overlay — SWE-bench Verified resolution 66.6% → 71.8%.

Industry Practitioners' Perspectives on AI Model Quality

EASE 2026 · invited to IEEE Software · CORE A

Mixed-methods study (15 interviews + 50-practitioner survey) of how industry actually manages AI quality — e.g., most practitioners never monitor deployed models for data drift, a critical blind spot for production ML.

Backdoors in Code Summarizers: How Bad Is It?

ASE 2025 · CORE A*

First systematic (9-factor) study of data-poisoning backdoors in code LLMs: just 20 of 454,451 samples (0.004%) implant a backdoor the standard spectral-signature defense fails to remove, and smaller training batches make the attack stronger — a concrete, under-estimated AI supply-chain risk.

Gotcha! This Model Uses My Code! Evaluating Membership Leakage Risks in Code Models

IEEE TSE 2024 · co-author

A membership-inference attack for code models that jointly uses model input, output, and ground truth to detect training-data membership.

Unveiling Memorization in Code Models

ICSE 2024 · CORE A* · co-author

Empirical study of how large code models memorize training data, with security and privacy implications.

What Do Users Ask in Open-Source AI Repositories? An Empirical Study of GitHub Issues

MSR 2023 · CORE A · co-author

A taxonomy of GitHub issues across 576 open-source AI repositories, characterizing what users ask and how issues get resolved.