Each project is its own repo and its own deploy. Some are live, some are works in progress, all of them taught me something.
Multi-service GCP system that triages support tickets by predicted priority with a fine-tuned DistilBERT classifier. FastAPI backend on Cloud Run with Firebase auth and per-user rate limits, Cloud SQL Postgres scoped by uid, Vertex AI for training and the model registry.
Implementing a transformer-based LLM in PyTorch from first principles, following Sebastian Raschka's Build a Large Language Model. A working understanding of every line, not just the API.
A Claude Code skill and matching sub-agent that turn the main agent into a coordinator. Decomposes a plan into waves, spawns implementer sub-agents in isolated git worktrees, runs them in parallel where safe, and stays out of the way between checkpoints.
U-Net CNN for image segmentation that automates cell counting for biomedical research. Mean error of 1.4 cells, beating the 3-cell target by 2x and replacing manual counting.
Full-stack AI wallpaper generator pairing a FastAPI backend with a React UI. Generates SDXL images, upscales with Real-ESRGAN to 4–5K, and ships with 15 device presets and a searchable gallery. WebSockets stream live progress.
An MCP server that lets Claude Code screenshot SVG files in-context — pass a path, get back a rendered PNG. Closes the visual feedback loop for SVG diagramming, so Claude can see what it's drawing without leaving the terminal.
A self-contained ~95KB single-file visual playground — generative gradients, lava lamps, starfields, liquid glass — with a layerable post-FX pipeline (bloom, film grain, vignette). All state round-trips through the URL hash so any look is shareable.
Polls recreation.gov and ReserveCalifornia for campsite availability on a fixed date range and pushes a notification the instant something opens up. Both providers expose unauthenticated JSON APIs — no scraping, no token, just diff-on-poll.