I build ML systems that survive contact with production.

A React + Node point-of-sale app for small retail — built to survive ghost-item bugs and Render's free-tier memory resets, not just to demo well once.

• Fixed deleted items silently reappearing after backend restarts
• Added a tallytap_local_catalog localStorage cache so the catalog survives Render cold starts
• Polished the layout specifically for laptop screens, where it's actually used

A lead-scoring platform built with two teammates — five ML models working together to tell sales teams which leads are actually worth calling.

• TF-IDF + feature hashing feeding sentiment, intent, purchase-intent, lead-score regression, and lead-temperature models
• FastAPI backend on Render with PostgreSQL/Supabase and Twilio for call handling
• Wrote a Hindi setup guide so the rest of the team could actually run it locally

A full translation ecosystem shipped in a single 6-hour sprint — one cloud API powering both a Windows desktop app and an Android app.

• Flask + deep-translator backend, served with Gunicorn and deployed on Render
• Same API wrapped into a Windows .exe (Python webview + PyInstaller) and an Android APK
• Translation runs entirely in the cloud — zero heavy dependencies on the client

A CatBoost + LightGBM ensemble that verifies how genuine a customer's interest is, from a 46-feature pipeline — with an API documented well enough that someone else could actually use it.

• Swagger-validated FastAPI with /analyze, /batch, and /feature-importance endpoints
• Ensemble tuned for the kind of messy, inconsistent data real customers produce

A Solidity smart contract for token staking and rewards — built to understand how DeFi mechanics actually work under the hood, not just how to use one.

• ERC-20-style token combined with on-chain staking and automated reward distribution
• Reward math is transparent and verifiable on-chain instead of trusted from a backend