I build the systems that decide which model gets used, and evaluate whether they should be trusted. My work sits between production AI engineering and applied ML research — LLM evaluation, agentic pipelines, and full-stack delivery.
An AI/ML/DS specialist with strong expertise across machine learning, NLP, and bioinformatics, working as an AI Engineer at MSD. My hands-on experience spans generative AI applications — agentic orchestration, LLM evaluation, applied AI research, and automation — blending modern generative AI techniques with traditional logic to build practical, data-driven solutions.
Currently pursuing a part-time Master of IT in Business (AI Track) at Singapore Management University, with an undergraduate background in Data Science & AI from NTU. I'm comfortable moving between research contexts (SMU coursework, comparative studies) and production contexts (MSD deployments, CI/CD, infrastructure-as-code) — and I try to keep both grounded in the same standard: results that hold up under scrutiny, not just look good in a slide.
Researching how to benchmark LLMs against public benchmarks and internal gold-standard datasets — determining which benchmarks matter to the company, and how to route the right model to the right use case. Part of the AI Safety, Security & Ethics group; evaluates LLM safety around harmful, biased, or unfair outputs, and assesses agent-specific risks (prompt injection, remote code execution, data protection).
RAG-based retrieval POC combining fuzzy TMX matching, NER, longest-common-substring glossary lookup, and TMX-over-glossary precedence. Extending toward a self-evolving loop that updates TMX from post-translation edits, with RLHF/DPO under investigation.
Extracted structured data from unstructured, scanned, handwritten, or typed patient intake forms. Compared traditional OCR, LLM-based OCR, and general-purpose LLMs to pick the best-performing method, then mapped outputs to the backend schema.
Comparative study of memory-augmentation, skill-library, and RL-based routing as alternatives to retraining a frozen base model — evaluated across DROP, MMLU, GAIA, and four others. Built the Magentic orchestration layer and ran the baseline evaluations that the three strategies are compared against.
Automated screening against the INTERPOL 2026 fugitive database, replacing manual similarity matching with auditable, tiered risk scores. Trained a custom Vision Transformer to produce disguise-robust image embeddings for reliable similarity search under adversarial appearance changes.