Work

Projects

A selection of production systems and open-source projects focused on applied machine learning and data analytics.

Demo

AI Judge — SCOTUS Verdict Backtester

Can AI predict Supreme Court rulings? Retrieves 200 landmark SCOTUS decisions from CourtListener and runs a RAG pipeline with temporal leakage prevention — Claude 3.5 Haiku on AWS Bedrock, FAISS vector search, MLflow experiment tracking, and a Streamlit dashboard. 45.7% backtested accuracy.

45.7% backtested accuracy Temporal leakage prevention 200 SCOTUS cases

Python AWS Bedrock Claude Haiku FAISS MLflow Streamlit SQLite

GitHub

Archived

Fate/RAG — RAG Chatbot (Archived)

Production-grade RAG chatbot for the Fate Series universe on AWS. Full serverless stack: FastAPI on Lambda via Mangum, API Gateway, OpenSearch Serverless k-NN vector search, and AWS Bedrock (Claude Sonnet + Titan Embeddings). GitHub Actions CI/CD pipeline runs pytest on every PR and auto-deploys to AWS via CDK on merge to main. Docker Compose for local dev.

Serverless infra k-NN vector search (HNSW) CI/CD auto-deploy SSE streaming

Python AWS Bedrock Claude Sonnet OpenSearch FastAPI AWS Lambda API Gateway AWS CDK GitHub Actions Docker pytest Mangum Pydantic

GitHub

Demo

Pokémon Red — RL Agent

Modular reinforcement learning codebase training a PPO agent to play Pokémon Red from pixel observations. CNN policy network (~1.7M params), parallel environment vectorisation, TensorBoard monitoring, and checkpoint management — all running locally.

Python PyTorch PPO Gymnasium PufferLib TensorBoard

GitHub

Demo

Super Bowl LX Prediction

XGBoost binary classifier trained on 25 years of NFL data to predict Super Bowl outcomes. Key finding: regular season wins contribute 0% feature importance — defense (37%) and offense (35%) dominate. Honest evaluation with GridSearchCV and 5-fold CV.

67.5% CV accuracy Defense #1 feature

Python XGBoost scikit-learn Pandas Jupyter AWS S3

GitHub

More projects in progress. Want to collaborate?

Get in Touch