I'm a Computer Science and Spanish major at USC. I also research how fruit fly synapses maintain stable neural function in the Dickman Lab and build software with Scope.
Lately I've been building Agent Receipt and Counterstep, two tools for keeping AI-agent work reviewable. Agent Receipt checks a completed run against the permissions it was given; Counterstep takes the next step, rechecking current state and repairing only the reversible parts it was allowed to change.
I'm also continuing to build Atrium, an iOS-first strength-training app that keeps reliable workout logs, tracks recovery, and helps decide what to train next. It works offline, restores interrupted sessions, and keeps any AI-suggested changes behind a review-and-apply step.
I also built Mechanism Canvas, a shared organic chemistry workspace where a student and an AI agent can work on the same reaction without handing over control. The student decides what help is allowed, and the app checks the chemistry before work is committed.
| Project | What it is | Built with |
|---|---|---|
| Nish | A comedy oracle that gives deliberately wrong answers to harmless questions, with safety checks and no stored conversations. | React, TypeScript, Cloudflare Workers AI |
| Atrium | Offline-first strength-training app for logging workouts, tracking recovery, and deciding what to train next. | React Native, TypeScript, SQLite, Supabase |
| Agent Receipt + Counterstep | Reviews completed AI-agent runs against their permissions, then repairs reversible oversteps within the same fixed boundary. | Next.js, TypeScript, IBM Granite, Google Cloud |
| Mechanism Canvas | Lets a chemistry student and an AI agent work on the same reaction while the student stays in control. | React, TypeScript, WebMCP |
| Autocomplete Engine | Blends a C++ trie and bigram model with a small PyTorch LSTM to suggest the next word in about 5 ms. | C++, Python, PyTorch |
| Open-source contributions | Fixes merged into Apache Airflow and the Kubernetes Python client. | Python |