Passion Projects

Semantic knowledge graph of the anime ecosystem in Neo4j with a formal OWL/RDF ontology, NLP-derived emotion edges, and a five-technique explainable recommendation engine served via FastAPI and a Next.js frontend.

Profiles a pretrained Real-ESRGAN checkpoint down to its GPU bottleneck and replaces it with a custom fused Triton kernel, validated with a 19-test correctness gate and an honest, statistically-significant interleaved benchmark.

Multimodal deep learning system that predicts discrete emotion labels and continuous valence/arousal scores from Japanese text and audio using a Mixture of Experts architecture with learned soft gating, BERT and Wav2Vec2 encoders, and a FastAPI backend with WebSocket streaming.

Predicts VC-backed startup survival using a Bronze–Silver–Gold pipeline, star-schema modelling, and LightGBM / XGBoost / GRU models tracked in MLflow. Live Streamlit app.

Predicts Goodreads book ratings from metadata and sparse NY Times bestseller signals — six MLflow-tracked models, SHAP explainability, and an R Shiny dashboard with a live what-if predictor.

Streams UK Railway data into Microsoft Fabric via SCD Type 2 ETL pipelines and surfaces delay and revenue KPIs in Power BI.

End-to-end computer vision pipeline that predicts coat length and grooming needs from photos of Mimi using a fine-tuned MobileNetV2 multi-output model, with a weather-aware Streamlit inference app.

Full-stack ML system that recommends recipes from available ingredients using TF-IDF vectorization, Truncated SVD, and KMeans clustering, with a FastAPI backend deployed on Render and an anime-inspired frontend on GitHub Pages.