CV

LLM Benchmark Auditing — Surface-Feature Leakage (TruthfulQA-476)

• Showed binary-choice 'truth' benchmarks leak surface features: a six-feature logistic regression (SURFACE6) separates correct from incorrect answers on TruthfulQA at 68.9% accuracy (AUC 0.714) — above the leading leaderboard model — and audited 15 benchmarks (HaluEval QA worst at AUC 0.973).
• Built Audit-Prune and released TruthfulQA-476, a cleaned subset that drops audit AUC to chance (0.528) while preserving model rankings (Spearman ρ = 0.915).
• Paper: Judging by the Cover: Cleaning Truth Benchmarks to Avoid Surface-Level Feature Leakage — under review.

LLM Activation Steering and Robustness under Dataset Corruption

• Controlled LLM verbosity and tone via residual-stream activation steering in Llama 3.1 and Qwen-2.5, with layer/strength sweeps for interpretability.
• Built a guardrail evaluation pipeline using LLM judges and JSON-schema validation, reducing hallucinations.
• Studied and mitigated noisy and adversarial corruption in steering datasets via robust high-dimensional mean estimation.
• Created and released three public contrast datasets on Hugging Face (conciseness–verbosity, positivity–negativity, formal–informal) for steering and evaluation.
• Paper: Understanding and Mitigating Dataset Corruption in LLM Steering — 2026.

Efficient Multi-Dimensional Two-Sample Testing (dKS)

• Extended the Kolmogorov–Smirnov distance to multidimensional probability distributions and robust two-sample testing.
• Implemented an ε-approximate algorithm with near-linear O(n log n) run time in 2D; unit-invariant integral probability metric (IPM) with ε-accurate grid method; stronger tests than unstable mdKS.
• Paper: Efficient and Stable Multi-Dimensional Kolmogorov–Smirnov Distance (arXiv: 2504.11299).

MotionPI — Privacy-First Wearable Sensing for Behavior Analytics (Full-Stack)

• Built an on-device wearable analytics stack (MotionSense HRV wristbands → app → API → database); streamed PPG/ENMO/survey with offline-first, schema-validated sync; sustained ~7.7M records/day with zero malformed writes.
• Built a visualization tool for tuning activity-trigger thresholds.
• Reported and verified a Flutter SDK bug (Issue #166937), later confirmed and triaged by the Flutter team.
• Paper: Designing a Secure and Resilient Distributed Smartphone Participant Data Collection System — EAI SmartSP 2025.

Anomaly Detection — Region-Aggregated Spatial Scan Statistics

• Replaced centroid scans with multi-point region sampling to increase detection power under axis-aligned rectangles; reproducible experiments with a C++ backend.
• Paper: Sampling for Region-Aggregated Spatial Scan Statistics — under review (double-blind).

Computer Vision — High-Dimensional Spectral–Spatial Change Detection

• Created a dual-stream 3D/2D CNN with SE attention; accuracy >96%, κ > 0.9, and lower false positives vs. baselines.
• Paper: A Hyperspectral Change Detection Framework Based on Double-Stream CNNs with Attention Module.

Ph.D. in Computer Science

• Advisor: Prof. Jeff Phillips.
• Expected graduation: May 2028.
• GPA: 3.9/4.0
• Focus: LLM interpretability; high-dimensional geometric data analysis.
• Relevant coursework
  • Machine Learning
  • Probabilistic Machine Learning
  • Deep Learning
  • Data Mining
  • Visualization

M.Sc. in Computer Science (Algorithms & Computation)

• GPA: 17.74/20
• Thesis: Graph-theoretic modeling of bushfire propagation.
• Relevant coursework
  • Advanced Algorithms
  • Approximation Algorithms
  • Randomized Algorithms
  • Quantum Algorithms & Computation
  • Graph Algorithms
  • Network Science
  • Internet Algorithms
  • Distributed Systems

Graduate Fellowship

Top-Talent Scholarship (merit)

M.Sc. University Entrance Exam — Rank 29/20,000 (top 0.15%)

Talk — Designing a Secure and Resilient Distributed Smartphone Participant Data Collection System

Poster — Efficient and Stable Multi-Dimensional Kolmogorov–Smirnov Distance (dKS)