Projects
Things I've built and worked on.
FacetForge
- Built an end-to-end CLI that transforms unstructured Shopify catalog data into queryable product facets using deterministic extraction, embeddings, reranking, LLM classification, and computer vision with confidence-based abstention.
- Designed a resumable, cost-aware inference pipeline with closed vocabularies, Zod schema validation, granular caching, concurrency controls, exponential backoff, provenance tracking, and offline provider fallbacks.
- Implemented automated taxonomy discovery, text-vision fusion, human correction workflows, Shopify metafield exports, and HTML/CSV coverage reports; validated the system with 325 deterministic tests across 25 files.
BikeShare
- Built a full-stack peer-to-peer micromobility marketplace that lets students list, search, and rent bikes, scooters, and skateboards through renter, owner, and admin flows.
- Implemented indexed geospatial search with PostGIS radius queries, supporting distance, price, vehicle type, rating, and availability filters with nearest-first ranking.
- Engineered privacy-preserving location logic that shows approximate pickup areas publicly and reveals exact coordinates only after owner acceptance.
- Designed overlap-safe booking logic using interval conflict checks and row-level locking to prevent double-booking during concurrent owner acceptances.
Poker AI Bot
- Implemented an LSTM-based poker agent trained via self-play reinforcement learning, achieving a sustained win rate of +30 mbb/1000 hands.
- Integrated the DeepSeek LLM API into the preflop evaluation pipeline, reducing computational overhead by 35% while maintaining 90% accuracy on preflop hand analysis.
Boiler Bites
- Built a Purdue dining hall macro tracker that lets users log meals, set macro goals, and view daily calorie and macro totals.
- Automated daily ingestion of 5+ Purdue dining courts via Vercel Cron, delivering 99%+ data freshness and supporting 1,000+ weekly meal logs hands-free.
- Hardened production with Sentry + logging, achieving 99.5% uptime and cutting failure debug time 70%+ at peak hours.