I build coding agents.
And lead the team
behind them.

I architected and prototyped dynamic, specialized subagents to split independent work into parallel tasks. Each works in its own context, with a shared scratchpad for progress—keeping the main thread focused and helping control token cost.

I led permission controls for tools and terminal commands, protected files and folders, and ways for agents to ask for clarification. Developers choose what an agent can do, and which decisions need their input.

I led six months of context engineering improvements in early 2026, raising internally measured task success from 82% to 90–95%. Prompt caching, pruning stale tool outputs, and summarization also delivered an internally estimated 3x improvement in token-cost efficiency. The work includes hybrid retrieval, reranking, repository maps, and file-based memory.

We use Langfuse and reasoning models as LLM judges to evaluate every new model and prompt improvement. We update the judge as our evaluation needs change.

Monthly paid-subscriber retention improved from 65% to 70% over eight months. Monthly enterprise customer retention was 90%, counting existing customers and excluding new sign-ups. These are internal product metrics.

I guide Cloud Agent architecture for repository-connected work in isolated execution environments. Checkpoints, model fallback, and validation-to-pull-request workflows give developers ways to review and recover agent changes.

Earlier at PureCode, I designed an image-to-code ML pipeline using YOLOv7-based UI detection and an intermediate DSL, built training-data tools, and engineered a transpiler for multiple frontend frameworks. I also helped build chat-based UI iteration and the subscription, payment, and workspace systems around the product.