Data Scientist & ML Engineer | Master of Data Science @ UBC | MBA Finance + B.Tech
I build data-driven solutions at the intersection of finance and machine learning.
An automated property attribute pre-filling pipeline using vision transformers and confidence calibration.
- Foundation Models: Fine-tuned DINOv2 and RemoteCLIP vision transformers using parameter- efficient unfreezing and Layer-Wise Learning Rate Decay (LLRD).
- Background Masking Pipeline: Integrated Microsoft Building Footprint and OpenStreetMap polygons with a custom U-Net segmentation fallback to prevent background shortcut learning.
- Underwriting Calibration: Implemented Platt scaling and Isotonic regression calibration to achieve 96.83% precision and 35.46% coverage at the 95% confidence threshold.
- Experiment Tracking: Logged and tracked hyperparameters, reliability curves, and UMAP projections using MLflow hosted on AWS (EC2/S3).
A production-grade, serverless MLOps system with 3 predictive strategies competing live on the S&P 500.
- Alpha-Driven Strategy Performance: Achieved a +18.8% Alpha over SPY (cumulative excess return) and a 1.54 Sharpe ratio over a 7-month live run (Oct 2025 – May 2026), outperforming SPY (+3.94%) with a -9.1% maximum drawdown limit.
- Multi-Model Machine Learning Engine: Evaluates three competing strategies: Fama-French momentum factor model, XGBoost classifier with 15 technical indicators (RSI, MACD, etc.), and an LSTM network utilizing 60-day price sequences across 503 tickers backtested on 9 years of historical data.
- Dockerized AWS MLOps Pipeline: Daily automated ETL pipelines and parallel inference scheduled via EventBridge and run on containerized AWS Lambda functions, persisting prediction logs to S3 and SQLite cache.
- Streamlit Dashboard & CI/CD: Live Streamlit application hosting real-time strategy returns, performance benchmark charts, and automated daily model health checks backed by automated GitHub Actions tests.
- Live Dashboard
Amazon Product Query Assistant — Hybrid RAG Search Assistant
An offline Retrieval-Augmented Generation (RAG) system for querying the Amazon Appliances reviews dataset.
- Hybrid Retrieval: Combined BM25 lexical search (rank-bm25) and semantic vector search using Sentence-Transformers (all-MiniLM-L6-v2) indexed via FAISS.
- Local Generation: Deployed a local Ollama LLM orchestration (Llama 3.2 3B) for generating grounded answers with citations.
- Data Engineering: Developed a fast data ingestion and preprocessing pipeline leveraging DuckDB to stream and convert raw reviews and metadata to Parquet format.
- Web Interface: Wrapped the retrieval and generation components in an interactive Streamlit UI.
Master of Data Science — University of British Columbia (2025-2026)
MBA Finance + B.Tech Computer Engineering — NMIMS University
- Website: pat0216.github.io
- LinkedIn: prabuddha-tamhane
Open to opportunities in Data Science, Machine Learning Engineering, Financial Analytics, and Business Analytics