Building small AI models that run where they shouldn't.
- Scribble detection model — optimized for CPU inference. No GPU, no cloud. The constraint is the research.
- AI orchestration engine — pay-as-you-go chat interface on OpenRouter. Model-selectable, no subscription lock-in. Shipped as a React Native app, published on Play Store.
| Area | What I actually did |
|---|---|
| Context management | Sliding window, token budgeting, stateful multi-turn |
| Multimodal | Image support for models that aren't natively multimodal |
| Sandboxed execution | Remote Python execution in Docker isolation |
| Web search | Tool-calling via OpenRouter, real-time retrieval |
| Mobile shipping | OTA vs store updates, 2–3 day review cycles, ES compat |
- Mainnet-deployed dApp — ecosystem grant-backed, unusual for a 3rd year CS student
- PR open in Uniswap v4-core
- Built Ghost, a privacy hook, for the Phoenix Buildathon
- RAG agent — pgvector retrieval, tool-calling, Docker-isolated Python execution
Python · React Native · TypeScript · Node.js · PostgreSQL · Solidity · Foundry
anxbrt.dev · Twitter · LinkedIn
3rd year CS @ SOA University · Open to applied AI internships at early-stage startups