Turned a capture-the-flag crypto eval into a GRPO training environment for Qwen3-4B on prime-rl. Kept an honest null result (no significant lift at n=16 rollouts), diagnosed it to a difficulty cliff that starved GRPO of gradient, then built a procedural difficulty generator that splits each task across decoupled sub-knobs to turn a cliff into a climbable curriculum. Published as a repo, a wandb report, HF weights, and a Prime Hub environment.

A language model that reads forward-English questions and answers in character-reversed English, generated tail-first. Custom BPE trained on a mixed forward + reversed corpus, direction tokens for mode control; standard decoder-only transformer, forked from nanochat.

Agentic system that turns any URL into a dev-ready multi-page Next.js clone. Playwright crawls and inspects the source site at desktop + mobile breakpoints; page-worker agents produce a structured Section IR; deterministic codegen writes a real Next.js App Router project. Follow-up chat edits target a single page/section and regenerate only that slice.

iOS app displaying dining hall menus across university campuses. Created with Swift/SwiftUI and AWS infrastructure for automatic menu updates.

Swift-based iOS app for automated receipt tracking, using Square Payment API and AWS. Built website for stores to download extension with NextJS and Tailwind.