Projects

A cloud for small software: a person or an AI agent deploys a Python or JavaScript program, and anyone with the link can run it without an account. Each program runs in its own WebAssembly sandbox with a time limit and reaches only what its owner granted. The research runtime it grew from is described in a paper that covers its architecture, analyzes which isolation properties the code enforced, and measures about 0.15 MiB of memory per added isolate.

Grafted a small recurrent reasoning module with learned halting into a frozen language model, training about 5% of the parameters. It took Pythia-1B from near 0% to 54% on eight-hop questions and 61% on entities it never saw in training, and reached 96% on Qwen2.5-Coder-7B. Also studying when a deployed model should learn a new fact in its weights or look it up: looking it up was as accurate or better on every test, and on Pythia-410M learning a fact only saves compute once it has been asked about roughly 2,400 times. Not yet published.

ESP32 sensor packs and a thermal camera measure sound, room conditions, and crowd movement and feed a Raspberry Pi. A 1.8M-parameter GRU trained on about 3,000 scraped tracklists predicts the next song, and a small reinforcement learning head adjusts each pick to how the crowd moves. Built with Chris Spletzer.

Built a generator that sets up a new bank's AWS infrastructure from one short YAML file, replacing about ten hand edits per bank. Built Studio, an MCP gateway that gives AI tools one way into the company's internal apps under one set of access rules, and rewrote it in TypeScript after security review. Shipped eleven changes to the Spring Boot portal banks log into, and built the pipelines behind self-updating documentation.

A therapy chatbot prototype built at New York Brain & Spine Surgery: clinical rules steer each conversation, and the model writes the replies. One system covers several conditions. The practice later canceled the project.

A mixed-architecture K3s cluster (two Raspberry Pi 4s, an Ubuntu iMac control plane, a Lima VM on an Intel MacBook) with encrypted secrets, tested snapshot restores, and monitoring over Tailscale. It runs long unattended jobs so the laptop does not have to.

A small control plane for renting GPU boxes on the Vast.ai marketplace for unattended multi-hour training: race several offers, health-probe each on real hardware before trusting it, run the job detached, sync checkpoints to S3, and destroy the box at a deadline that survives the laptop closing. The event log of 791 create calls and 360 probe verdicts over two months shows where rentals actually fail: 34% of creates, 29% of probes on SSH, and only 4% on genuinely bad hosts.

A full-stack options-trading sim with 61M financial data points on AWS. Users build and backtest strategies and watch them play out against historical market data.

A search engine for ~23,000 Dropbox documents, with sentence-level highlight extraction. Partitioned indexing + Dropbox API integration for real-time access. Lehigh capstone for an early-stage startup.