I make agentic coding reliable, repeatable, and reviewable.
I build an agentic coding platform that standardizes how engineers run AI coding agents. It turns one task into governed workflow steps, runs each with a real coding CLI, and puts a deterministic security gate on every action. Prompt-and-pray → versioned, reproducible engineering.
Right now I work full-time at the frontier of AI-assisted development: how do you take agentic coding from a one-off prompt to a repeatable, reviewable workflow an engineering org can depend on?
I design and build an agentic coding platform — a desktop app that turns a high-level task into governed workflow steps, each executed by a real coding CLI (Claude, Codex, Gemini and more) spawned as a background agent. A deterministic governance gate sits between every step and the code it touches, and a tamper-evident audit log records each decision. The hard part was never the model. It's the operations around it.
↑ a faithful simulation of the platform's control plane — task queue, workflow steps, live agent runs, streaming traces.
I implemented a system that turns broad AI requests into ordered, trackable execution steps. Each step runs with clear state, progress visibility, and separation between planning, implementation, review, and follow-up work.
I built a control layer that checks AI actions before they continue, including approval rules, sensitive-data handling, and audit visibility. This makes AI-assisted work easier to trust, review, and operate in a real engineering environment.
I created a context approach that gives AI the right project knowledge before it acts. By supplying focused project structure, relevant files, and reusable instructions, the system helps produce more consistent and repeatable results.
Side projects built the way I build at work: agentic coding from a written plan, with review before anything ships. Each one is a single native binary on GPUI, with no Electron and no web view, shipping for macOS, Windows and Linux.
Four arcade classics, each built as a small state machine — game loop, collision, enemy behavior, scoring, and browser-saved high scores running on a raw <canvas>, no engine. Same discipline as the day job, just lower stakes and more fun. Drop a coin in JEREMIE ARCADE: grab the joystick and FIRE button, or use arrow keys + space.
let's make AI coding something a team can actually depend on