Data Scientist / ML Engineer Intern at Starhub Ltd (2025-05 – 2025-11)
Improved chatbot tool-calling accuracy, built agentic regression test harness, engineered MILP optimization engine, and reduced geomobility feature drift.
- Improved Agent Atlas chatbot tool-calling accuracy by 30%+ by integrating Model Context Protocol (MCP) and refactoring tool schemas/response parsing to reduce malformed calls on complex customer queries
- Built an agentic regression test harness (LangGraph/LangChain) for Agent Atlas that generates and validates 10+ synthetic tool-call test cases per run, catching tool-output regressions before release via automated assertions and benchmarks
- Engineered and deployed a Mixed Integer Linear Programming (MILP) optimisation engine for the Smart Retail Platform using OR-Tools and Apache Airflow; Utilised TelCo data to inform client decisions on advertising asset placements, achieving globally optimal results
- Reduced geomobility feature drift in a user home and workplace location classification pipeline by 65% (distribution error, measured by Population Stability Index (PSI)), by diagnosing data migration issues, retraining and validating fixes against previous valid runs, restoring reliable insights for dashboards
Lab Tutor at National University of Singapore (2025-01 – 2025-05)
Conducted weekly learning sessions on Java programming and automated grading for assignments.
- Conducted weekly 2-hour learning sessions for 30+ students on the Java programming language, focusing on programming paradigms such as Object Oriented Programming and Functional Programming
- Cut grading turnaround for 150+ Java assignments by automating compilation, test execution, and result summarization with Bash scripts and deterministic test fixtures
AI Engineer Intern at Pt. Adi Citra Sakti (2024-12 – 2025-01)
Trained and containerized facial recognition model and built evaluation framework for multi-class multi-label model.
- Trained and containerised a facial recognition model capable of processing multiple video streams using OpenCV, Docker and image augmentation techniques, achieving a precision of 70% via hyperparameter tuning, ensuring accurate identification across concurrent real-time inputs
- Built a standardised evaluation framework utilising confusion matrices to audit the multi-class multi-label model, streamlining the testing pipeline to reduce manual analysis time, ensuring robust validation prior to deployment
