Chua Jia Yang

Data scientist building production ML & GenAI systems. Pursuing an M.S. in CS at Georgia Tech.

Singapore · Email · GitHub

About

I'm a data scientist who likes building things that ship — models that wake up every morning and do useful work for real users. My interests sit across production ML at scale, reinforcement learning, and applied LLMs.

Outside of work, I'm doing my M.S. in Computer Science at Georgia Tech, build side projects on the weekend, and unwind with chess (currently around 1800 on chess.com) and basketball.

Work

DBS Bank · Associate, Data Scientist · 2022 — Present

Currently working on ML and GenAI applications for consumer banking. I build, train, and maintain a portfolio of propensity models that identify which customers are most likely to take up the bank's products — surfacing hundreds of thousands of leads daily so marketing teams can send the right offers to the right people. I continuously evolve the pipeline with new models, features, and capabilities, and also work on applied generative AI tooling alongside the core ML stack.

Education

Georgia Institute of Technology · M.S. Computer Science · 2025 — 2027

  • CS6200 Introduction to Operating Systems

    Systems-level course on processes, threads, synchronization, IPC, and distributed services — implemented in C.

    Key projects: Multithreaded GETFILE server (boss–worker pool) · Distributed file transfer over gRPC · Inter-process cache via shared memory

  • CS6250 Computer Networks

    Survey of modern networking — from link-layer protocols up through SDN, BGP, and internet measurement.

    Key projects: Spanning Tree & Distance Vector routing · SDN with Mininet/POX · BGP hijacking & measurement

  • CS6601 Artificial Intelligence

    A foundations-of-AI course spanning classical search, game playing, probabilistic reasoning, and machine learning.

    Key projects: Adversarial game agent (Isolation) · Gaussian Mixture Models with Expectation–Maximization · Hidden Markov Models with Viterbi

  • CS7638 AI for Robotics

    Probabilistic robotics — localization, control, and planning, framed around real robot problems.

    Key projects: Kalman Filter · Particle Filter · SLAM (Indiana Drones)

  • CS7642 Reinforcement Learning & Decision Making

    Theory and practice of RL — from tabular methods through deep and multi-agent RL.

    Key projects: Q-learning & SARSA foundations · DDPG on Lunar Lander · MAPPO multi-agent on Overcooked

  • CS7643 Deep Learning

    Modern deep learning — convolutional and recurrent architectures, transformers, and generative models.

    Key projects: CNNs from scratch · NLP with RNNs & Transformers · Generative models (VAEs & GANs)

  • CS8001 Agentic AI Seminar

    Hands-on seminar on modern LLM patterns — from prompting and RAG through to multi-agent systems.

    Key projects: Retrieval-Augmented Generation · Agentic frameworks (LangGraph, CrewAI) · Multimodal LLMs

  • Seminar Object-Oriented Programming with Java

    A seminar to pick up the fundamentals of Java — classes, inheritance, polymorphism, generics, exceptions, and JavaFX.

Singapore Management University · B.Sc. Economics (Hons) & B.BM Finance (Hons) · 2018 — 2022

  • Major Data Science & Analytics + Finance

    Graduated Summa Cum Laude. Dean's List 2018–2021.

Projects

A handful of side projects and coursework. All on GitHub.

Deep Learning · Computer Vision
GitHub ↗

Image dehazing: FFA-Net vs Diffusion

An interactive web tool for comparing two image-dehazing approaches side-by-side, trained and evaluated on aerial imagery from the DOTA dataset.

  • FFA-Net — Feature Fusion Attention Network for single-image dehazing
  • Diffusion Net — DDPM-based generative dehazing approach
  • Flask backend serving model inference; HTML/JS frontend for live image upload & comparison
  • Built as part of Georgia Tech Deep Learning coursework
PyTorchFFA-NetDDPMFlask
Reinforcement Learning
GitHub ↗

Lunar Lander · DDPG from scratch

A Deep Deterministic Policy Gradient agent implemented from scratch to solve the LunarLanderContinuous environment from Gymnasium.

  • Actor–Critic networks for continuous action spaces
  • Target networks with soft updates for training stability
  • Experience replay to break correlations in sequential data
  • Exploration noise, TensorBoard logging, periodic eval & checkpointing
PyTorchDDPGGymnasiumTensorBoard
Reinforcement Learning
GitHub ↗

Super Mario · PPO with CNN policy

A PPO agent that learns to clear Super Mario Bros levels directly from raw pixel input.

  • Proximal Policy Optimization via Stable-Baselines3
  • CNN policy network processing stacked visual frames end-to-end
  • Custom Gym wrappers for frame stacking, action shaping, reward design
  • Eval harness + episode recording for visualizing learned behavior
Stable-Baselines3PPOgym-super-mario-bros
GenAI & Agents
GitHub ↗

Werewolf played by LLM agents

A multi-agent simulation of the Werewolf social deduction game where LLM-powered villagers and wolves reason, deceive, and coordinate — with no hardcoded strategy.

  • Structured reasoning chains — agents follow ANALYZE → FORM_BELIEF → PLAN steps each turn
  • Hidden-information coordination — wolves share intent at night, villagers must infer roles from behavior
  • LangChain orchestration over Gemini, with full game-log replay
  • LLM-powered game summarizer that turns logs into readable narratives
LangChainGeminiMulti-agentPython
MLOps
GitHub ↗

Titanic ML · FastAPI service

An end-to-end MLE project taking a model from training to a deployable, containerized REST API with CI.

  • scikit-learn RandomForestClassifier trained & serialized for serving
  • FastAPI + Uvicorn with typed request/response schemas
  • Dockerized for portable deployment to Railway / Render
  • GitHub Actions CI runs tests & builds the image on every push
FastAPIscikit-learnDockerGitHub Actions

Chess trainer

A browser-based chess trainer that lets you play against a Stockfish engine of adjustable strength, with real-time blunder detection.

  • Stockfish.js engine running in-browser, Elo configurable from 600–2300
  • Real-time blunder detection with suggested better moves & position evaluation
  • Chessboard.js for interactive UI; fully responsive desktop & mobile
  • Live demo deployed via GitHub Pages
JavaScriptStockfish.jsChessboard.js

Coffee vouchers

A tiny single-page web app for tracking the redemption of 10 digital coffee vouchers — built as a quick utility for friends.

  • Vanilla JavaScript frontend, no build step
  • Firebase Firestore backend with security rules locked down to redemption-only
  • Hosted on GitHub Pages
JavaScriptFirebaseFirestore

Contact

Always happy to chat about ML, RL, GenAI, or any interesting problem. The quickest way to reach me is email, or find me on GitHub.