AI & Data Engineering Intern at QuantumTX

Developed an 8-stage Python ML pipeline on 1,716 hospital patient records using XGBoost/gradient boosting with imputation and SHAP; achieved AUC-ROC 0.97 for dropout risk and 0.63 for treatment response. Migrated from Parquet to a production FastAPI + PostgreSQL + pgvector backend with longitudinal per-patient session tracking, idempotent concurrency-safe ingestion, and PHI-compliant processing. Built a config-driven ingest/clean/phenotype/outcomes/model pipeline from clinical Excel to Parquet/PostgreSQL, including a rules-based phenotyping engine generating 25+ binary flags across 14 condition groups with automated test coverage. Implemented a per-patient AI reasoning layer using Claude and embeddings stored in pgvector, plus a trend engine and Next.js dashboard with PDF report export.

AI Intern at Theia Health Pte. Ltd

Re-implemented SC-Net (MICCAI 2024), a DETR-style dual-branch model for simultaneous stenosis severity and plaque composition classification from coronary CT angiography volumes. Trained on a proprietary multi-cohort dataset of 797 patients across two hospitals; achieved cross-hospital performance with Stenosis F1 = 0.851 and AUC = 0.892, improving over the prior baseline. Designed a constrained 3D calibration system with per-class recall constraints to recover a clinically critical class; debugged silent architectural defects and built training infrastructure including loss engineering and distributed multi-GPU training in PyTorch.

Software Engineer at NUS IT

Built a cross-platform React Native health app with 28+ screens and Redux Toolkit, unifying iOS/Android and shipping via Expo EAS. Engineered a Node.js/Express REST API and a Python FastAPI AI microservice backed by PostgreSQL on AWS RDS with Redis caching and S3 presigned URLs; containerized with Docker and deployed on AWS EC2. Developed an AI coaching chatbot using AWS Bedrock and LangChain with persistent conversation history in PostgreSQL, plus a JITAI behavioral-intervention engine with adaptive trigger types. Architected a multimodal food-analysis pipeline using Claude Vision and a HuggingFace classifier, secured routes with JWT, OAuth, and role-based access control.