The lab

Stuart Labs

Stuart Labs is where my technical work lives. Most of it is infrastructure for agents, including MCP servers, harnesses and the tools I use to run several agents at once. The code is public where I can make it public, and every project has a write-up of how it works. My case studies and essays are here too.

Cornell Generative AICase studySpring 2025 – present

Running Cornell Generative AI like a startup

As president of Cornell's student AI engineering club, I rebuilt how we hire, delivered client work for two finance firms, led the move to our own products, and built the club's operations platform with coding agents.

Read the write-up →

Stuart LabsOpen sourceJune 2026 to present

ego-OS

A memory graph, stored on my own machine, that turns my notes, meetings, email and agent planning docs into memories, entities and concepts, and serves them to any AI agent over MCP. Every retrieval change is benchmarked, and one benchmark led me to delete a feature I had just built.

Stuart LabsProductAug–Oct 2026, ongoing

Outreach CRM

A multi-tenant CRM for student networking outreach. AI drafts each message from facts on file, the student sends it, and the whole pipeline doubles as an OAuth-protected MCP server that their own agent can work.

Stuart LabsOpen source2025–2026

HomeAssist and Tool-Calling Mini

An always-on voice assistant that I rebuilt from a Raspberry Pi prototype into an interruptible, proactive agent on the OpenAI Realtime API. Tool-Calling Mini is my QLoRA fine-tune of Qwen3-4B that can take over its tool routing from a self-hosted server.

Stuart LabsOpen sourceOctober 2026

echo-OS

A macOS menu bar tool that speaks a short headline when a long Claude Code turn finishes, tagged with the project folder and a per-session call sign, so I can leave several agents running and hear which one is done.

Stuart LabsOpen sourceSpring 2026

AgenticNodes

A local graph editor and runtime for agent workflows. You wire model, tool and data nodes on a canvas, run the graph, and inspect every node and edge event in per-agent swimlanes.

InstaLILYEssaySummer 2026

Build the Environment, Then Optimize the Harness

InstaLILY's two-part report on coding agents argues for mining merged PRs into a resettable environment and measuring its noise before letting an optimizer tune the harness. I led the design of that environment, Lily-Eval, and the evaluation work in Part 1.

Cornell CS 4701Open sourceFall 2026

HIVE: an eval-gated loop for improving agent harnesses

A Cornell CS 4701 project that freezes the model and lets a Claude Code overseer propose harness edits, kept only when they beat the current best on a graded suite. The scaffold runs end to end; so far it has run dry runs, not experiments.