projects

An iOS app where Marcus Aurelius, Seneca, and Epictetus respond to your evening journal entry. A solo project.

• Released on the App Store in July 2026.
• A crisis classifier screens each entry first and returns real resources instead of an AI reply.

• React Native
• Expo
• Hono
• Bun
• Supabase
• Claude
• RevenueCat

A demo airline support agent built on Sierra's open τ-bench domain, with a workbench that shows every tool call. An experiment in making agents dependable.

• Scored 0.800 pass^1 on a five-task slice. The methodology is in the repo.
• Refuses off-policy requests and cites the policy when it does.

• Claude Sonnet 4.5
• Vercel AI SDK
• Next.js
• Upstash
• Zod

A demo agent for a made-up New Jersey insurer. It quotes new customers and services existing ones, using the same pricing engine as a voice quote line I made.

• Passed 36 of 36 trials across 12 tasks (pass^3 of 1.000), graded on the final record.
• Verification and explicit consent are enforced in code.

• Claude Sonnet 4.5
• Vercel AI SDK
• Next.js
• Upstash
• Zod

An open source harness that runs state-verified tasks against any chat model and reports pass@k and pass^k. Tasks are graded on the final database state.

• Claude Haiku passed 20 of 20 trials. Llama 3.1 8B passed 12 of 20.
• Caught a small model reporting a cancellation it never made.

• TypeScript
• Bun
• Ollama
• Claude

An internal tool that replaced a spreadsheet for tracking litigated claims, with workflow checks and dashboards.

• In production since September 2026.
• Calculates the litigation gap automatically.

• React
• FastAPI
• SQLite
• Plotly
• Docker

A prototype that reads 300 to 500 page medical demand packets, rebuilds the injury timeline, and flags pre-existing conditions for an adjuster.

• Picks up new packets from an inbox and emails back a discrepancy summary.
• Fast models triage and extract, and a stronger model looks for discrepancies.

• TypeScript
• Bun
• Claude
• Gmail API
• Google Docs API

A fork of Daniel Miessler's PAI that adds skills, hooks, memory, and a verification loop to Claude Code. I use it for most of what's on this page.

• Hooks check every tool call before it runs.
• Tasks are checked against written criteria before they count as done.

• TypeScript
• Bun
• Claude Code
• MCP

An experiment that reads my Garmin data, tracks training load, and adjusts the plan. A Telegram bot sends a morning briefing.

• Uses TRIMP and the acute-to-chronic workload ratio.
• Paused for now.

• TypeScript
• Bun
• Garmin
• Telegram
• Claude

My friend Mike and I organized and ran a backyard ultra in Queenstown, Maryland. Every hour you run a 4-mile loop, and the last runner left wins.

• 52 people gave $4,310. After GoFundMe fees, $4,169.02 went to Tunnel to Towers and PGA HOPE.
• I ran 42.8 miles over ten loops. Mike outlasted me by one loop.

• Running
• Fundraising
• Event planning

A data pipeline for university researchers, built by a six-person team over a year. I was the product owner.

• A Raspberry Pi pipeline and C++ dashboard that write sensor data into MariaDB.

• C++
• Raspberry Pi
• MariaDB

A Streamlit app from my Johnson & Johnson internship that summarizes medical text at five reading levels, from beginner to specialist.

• Python
• Streamlit
• OpenAI
• LangChain

Minesweeper in Java and JavaFX with four difficulty levels, made with a four-person team for CSCI 205. I was later a TA for the course.

• Java
• JavaFX
• Gradle
• JUnit