Behavior of Chain-of-thought Monitorability
Studying how monitor effectiveness changes as the capability gap between monitor and target models widens, with a case study on distinguishing sandbagging from genuine incapability.
Projects
A collection of projects I've worked on, ranging from academic research to personal explorations. If you think it would be useful or interesting to collaborate on a project, please contact me to discuss.
Featured Projects
SEATauBench
Extended Tau2-Bench for low-resource Southeast Asian languages and localized agentic evaluation.
Mnemonic Generation for Vocabulary Learning
AI chatbot generating memorable English and Mandarin mnemonics with QLoRA fine-tuning and DPO preference modeling.
Mini-LLaMA2 PyTorch Implementation
Minimal Llama-2 implementation in PyTorch with RoPE, RMSNorm, SwiGLU, self-attention, and transformer blocks.
Other Projects
BenchFlow (Open Source)
Contributed to RL runtime environments and community-based AI benchmarks.
Multicultural Riddles Benchmarking
Analysis of LLM factuality, hallucination, topic patterns, and cross-cultural reasoning failures on multicultural riddles.
SEACrowd Website (Design & Growth)
Designed the SEACrowd website and managed social content for a Southeast Asian AI research community.
Astro Scholar Theme for Academics
Created and maintain a personal academic Astro theme for research portfolios, publications, projects, and blogs.
Animals-Aligned Agentic AI Benchmarking
Evaluated whether implicit values in agentic LLMs are inclusive of animals, by extending CaML's The Animal Compassion
Multilingual Representations with Sparse Autoencoders
Investigated multilingual LLM representations with SAEs and feature steering across 67 languages.
San Francisco Safety Dashboard
AI-powered Hex dashboard scoring safety across 41 San Francisco neighborhoods with live DataSF and ClickHouse data.
SportConnect
Luma-like platform for discovering, managing, and RSVPing to local recreational sport events.
SeizureSavvy
Full-stack PWA for seizure tracking and predictive warnings, with React, Flask, XGBoost, and LSTM models.
Ask My Second Brain with RAG
Question-answering system over Obsidian-style personal notes using LlamaIndex, OpenAI API, and retrieval-augmented generation.
Causal Inference of Political Intervention
Replicated and extended a synthetic control analysis of Philadelphia SNAP benefit redemption in R.
Bayesian Hierarchical Modeling for GP Visit Count Data
Analyze GP visit patterns using Zero-Inflated Poisson models with complete and partial pooling, with Bayesian inference and data imputation.
Text classification using SVM and Naive Bayes
Text classification system to automatically classify notes and assignments using SVM and Naive Bayes with 87% accuracy.
GitHub Activity
My GitHub ↗ contributions over the past year. Colored squares represent days with commits.