Retention analytics
Retail Customer Churn Prediction and Retention Analytics
Built an end-to-end churn system that connects feature engineering, warehouse modeling, explainable machine learning, and a Streamlit dashboard designed for retention action.
- Compared Logistic Regression, Random Forest, and XGBoost across 25,000 customers and 300,000+ behavioral events.
- Modeled 250,000+ transactions in Snowflake with dbt using a dimensional star schema for RFM, cohorts, and revenue-at-risk reporting.
- Used SHAP to expose top churn drivers and tied the workflow to projected annual savings of $1M-$3.5M.
Data product build
BaiBeta Climbing Shoe Price Tracker
Built a full-stack tracking product that ingests Canadian climbing shoe pricing daily and surfaces actionable deals through search, trends, and watchlist alerts.
- Built and managed a Supabase-backed pricing pipeline in TypeScript, using Prisma models to normalize daily store data and persist historical snapshots for trend analysis.
- Implemented a Next.js App Router experience with filters for brand, store, discount, size, and stock, plus Recharts-based history views so users can validate whether a sale is genuinely attractive.
- Applied CI/CD practices with automated quality checks and deployment workflows, while shipping Auth.js login, watchlists, and deduplicated email alerts to improve reliability and user retention.
Reinforcement learning
Deep Q-Learning Snake AI
Built a Deep Q-Network agent in PyTorch to learn Snake through self-play, with an end-to-end training loop, replay memory, and performance tracking.
- Engineered an 11-feature state representation for danger detection, direction context, and food position to drive action decisions.
- Implemented experience replay with an epsilon-greedy policy to stabilize learning and improve exploration-to-exploitation balance.
- Trained and evaluated the agent across hundreds of games, showing clear score progression and stronger collision-avoidance behavior.
Applied AI build
Cerebro-AI
Built a local study assistant around retrieval-augmented generation so a user can turn their own material into flashcards, practice prompts, and guided review.
- Designed a RAG workflow for flashcards, example questions, and content discussion.
- Connected a Google Drive knowledge base to a vector store for efficient retrieval.
- Integrated DeepSeek V3 API support while keeping the system grounded in user-provided study material.
