Projects
What I shipped, with the numbers and the caveats.
Code search for AI coding agents, so they stop guessing
Code-search platform for AI coding agents: hybrid BM25 + vector retrieval with reranking, tree-sitter AST chunking, a web app, a REST API and a 12-tool MCP server.
- 94% Hit@1, research eval
- 641ms p50 cold search, public playground
A QA analyst for voice agents. It listens to every call so you don't have to
AI QA analyst for voice agents: grades every call a Retell scheduling agent takes. Code checks what the call log proves, and an LLM-as-a-judge checks what the agent said against the clinic's real slots.
- 23/23 Planted failures caught
- 0.89 Macro-F1
Teach a pharmacy-finder to call less and find more
Active-sensing pharmacy search: predict which pharmacy most likely has a hard-to-find drug, rank who a voice agent should call first, and learn from every answer. Tested on a simulated pharmacist with real pharmacy data.
- 47% Fewer calls to find, simulated
- 10% to 0% False 'in stock' answers, simulated
A Claude.ai token and cost meter that runs entirely in the browser
Chrome extension that tracks Claude.ai token usage and cost in real time, entirely in the browser. Published on the Chrome Web Store.
- 1,808 Vitest tests
- Published Chrome Web Store
You're looking at it right now
A desktop operating system in a browser tab: window manager, terminal and dock, built from scratch.
- From scratch Window manager
- Mobile Phone shell
Writes the context files coding agents need
Statically analyzes a codebase and generates agent context files (AGENTS.md, CLAUDE.md, .cursorrules). Published on PyPI.
- 22 Releases on PyPI
- 3 Context formats
Because manual bookkeeping is so 2019
AI expense tracker with automatic categorization, category budgets, spending insights and a daily safe-to-spend figure.
- Auto Expense categorization
- Daily Safe-to-spend figure
Infrastructure that actually stays up
You know what's harder than writing code? Making sure it stays running in production. SecureScale is a fully automated AWS infrastructure setup designed to not break at 3 AM.
- 99.9% Uptime
- 2h to 10m Environment provisioning
Hackathon: auto-generate Journey Success test cases for IBM watsonx agents
IBM watsonx Orchestrate hackathon. The problem: manually authoring Journey Success test cases for watsonx agents is tedious and error-prone. Test Forge is a multi-tool agent that reads deployed agent specs and generates validated test cases automatically.
- 4 Tools implemented
- 13 Unit tests passing
Does an agent pick the wrong tool when too many MCP servers are loaded?
When an agent has dozens of MCP tools loaded, does it start picking the wrong one? Tool Crowding is a harness built to measure that cleanly, with the design and its hypotheses pre-registered before any data.
- 345 Harness tests
- 199 Synthetic tools
Serverless ingestion on GCP with multi-tenant isolation and PII redaction
Mem Machines is a serverless data ingestion pipeline built entirely on GCP: Cloud Run, Pub/Sub and Firestore working together on high-throughput data streams.
- 3 GCP services
- 0 Servers managed
Helping students find what they actually need
A clean, searchable directory of campus resources built to help students navigate the overwhelming maze of university services, tools, and support systems.
- Live At campus-resources.vercel.app
- Searchable Resource directory