agent runtimes · workflow engineering · developer tools · applied research
I started with web interfaces and build tooling. These days I work closer to the boundaries between agents, runtimes, data, and people—mostly in TypeScript, with Rust and Python where the system calls for them.
I care about making AI work operational: explicit state, recoverable execution, human checkpoints, useful observability, and evidence that the real path works.
| Project | What I am exploring |
|---|---|
| Agent Skills | Portable, evidence-first workflows for coding agents—designed around clear outcomes, adaptive depth, and truthful completion. |
| Ankole | Contributing runtime, console, and cross-runtime E2E work to an open-source, self-hosted AgentOS for shared AI colleagues. |
| MDX Renderer | A browser-based MDX playground with isolated preview, custom components, local persistence, and a deliberately small runtime. |
| agent-cli | An experiment in multi-agent software workflows with durable state, DAG scheduling, human checkpoints, and breakpoint recovery. |
inspect reality → model state & failure → ship a focused slice → verify the boundary → document the contract
- Real runtime evidence beats plausible-looking output.
- State, failure, recovery, and cancellation are product work—not cleanup.
- Tests, docs, config, and examples should ship with the behavior they describe.
- Repetition is usually a signal to build a tool, adapter, or reusable workflow.
- I prefer small, descriptive commits and progressive hardening over one opaque rewrite.
- Languages: TypeScript · Rust · Python · SQL
- Frontend: React · Next.js · Vite · Tailwind CSS · shadcn/ui · React-Native · Flutter
- Backend & data: Bun · Node.js · Hono · Drizzle ORM · PostgreSQL · ClickHouse · Redis · BullMQ · Polars
- AI & interfaces: AI SDK · MCP · oRPC · TanStack Query · Zod · hermes-agent
- Engineering: Turborepo · Biome · Vitest · Bun Test · Rstest · Docker · Kubernetes · Helm · OpenTelemetry · GitHub Actions
If you are working on durable agents, developer infrastructure, or better human–agent workflows, I am interested in the hard parts: ownership, recovery, and proving that the system actually works.