At 3M I built platforms for R&D, led migrations of fluid modelling applications to AWS Batch. Some fullstack work as well...
Check it out at joshupadhyay
Extemp – Impromptu speaking coach. Record a 2-minute speech, get structured feedback on clarity, filler words, and delivery. Built with Groq (Whisper transcription) + Next.js.
What's the Busiest Citibike Station? – 20 GB NYC Citibike dataset processed in 137s using Modal's parallel fan-out. An exploration of serverless GPU compute for large-scale data work, and how easy it is with Modal's functionality. Deployed using modal serve.
Fine-tuning Stable Diffusion – Fine-tuned SDXL on Edward Hopper's painting style. Covers the full pipeline: training, cold-start optimization, and serving a custom image model from Huggingface. Explores caching, GPU snapshots, and other optimization techniques.
NYC Gaussian – 3D reconstruction of St. Mark's Place using gaussian splatting algorithms on Modal GPUs. Converts Street View imagery into an explorable 3D scene in the browser. I generated my own Gaussians with AnySplat, and the PoC is available here.
The Crunch – AI-powered night-out planner for NYC. Claude tool calling + Mapbox for real-time venue recommendations, deployed on AWS EC2 with CI/CD via GitHub Actions.
Interior Bot (in progress) – Fine-tuning Qwen 2.5 7B on interior design using expert trajectories distilled from Claude. A ReAct agent + ChromaDB vector search, inspired by Databricks' KARL paper — small model, domain-specific retrieval, frontier-model quality at a fraction of the cost.
AWS (Lambda · ECS · Batch · CDK) · Docker · TypeScript · Python · Modal · GitHub Actions · Next.js