Tyson Thomas

Sheet 02 · About

I build machines that do the work.

I’m Tyson Thomas, an engineer who can’t stop building. I’m drawn to AI, robotics and software, and lately to the place where they meet: agents that leave the chat window and actually get things done.

Some of mine live in a browser. One lives on my desk, has antennas, and nods when it talks.

This site is the build log: what I made, what broke, and what I learned fixing it.

Engineer
AI · Robotics · Software
Building now
Agents, and a robot to run one
Find me
GitHub ↗X ↗RSS ↗

What I build

  1. P-01AI agent browser

    Browser Operator

    A web browser with an AI agent at the wheel, built on Chromium. You give it the goal; the agent does the browsing.

    • Chromium
    • Agents
  2. P-02Robot assistant

    Reachy Mini desk robot

    A small open-source robot turned desk assistant. Ask it to research something or plan a trip and it goes off and does it, nodding along while it talks. Its brain went from fully local, to OpenClaw, to Meta's Muse.

    A technical drawing on cream grid paper. On the left, Reachy Mini Lite, drawn from Pollen Robotics' open 3D model, with callouts for its body (head, antennas and base all move), ears (microphones on top of the head) and voice (a 5 W speaker in the body). Two arrows link the robot's head to a large shaded cube on the right labelled Muse, Meta's personal agent: "what you ask" goes to Muse, and "answers, sentence by sentence, + moves" come back, over the Muse gadget SDK. Four small cubes beside Muse are its sub-agents, which work in the background: research, trip planning, email + calendar, and lists + reminders. A title block above the link reads: Reachy Mini + Muse.
    • Robotics
    • Muse
    • OpenClaw
  3. P-03Hackathon build

    Reachy Mini × OpenClaw

    Two AI agents as teammates: Reachy Mini is the one in front of you, with voice, movement and personality, and OpenClaw is the one behind the scenes, reading links, summarising and remembering. Tell the robot "I'm sending an article now", drop the link in Discord, and the conversation just carries on. Built at the AI Tinkerers OpenClaw Global Unhackathon, February 2026.

    A technical drawing on cream grid paper. On the left, Reachy Mini, with callouts for its body (moves, nods, dances, shows emotion), ears (you talk to it face-to-face) and voice (OpenAI Realtime, gpt-realtime-1.5). Two arrows run through a box labelled Bridge API, port 8100, to a large shaded cube on the right labelled OpenClaw, personal AI assistant: voice transcripts and tasks go to OpenClaw, and findings and notifications come back for Reachy to say. Four small cubes beside OpenClaw are its tools, which work in the background: link research, browser tools, memory, and Discord plus other channels. A title block reads: Reachy Mini + OpenClaw. In front of you: Reachy. Behind the scenes: OpenClaw. Built at the AI Tinkerers OpenClaw Unhackathon.
    • Hackathon
    • OpenClaw
    • Reachy Mini
    Tech write-up ↗
  4. P-04Agent for real estate

    Realtor Agent

    An agent that works the real estate desk. It analyses properties, runs comparables and helps with the contract paperwork.

    A technical drawing on cream grid paper, labelled Fig. 01, one manager, three specialists. On the left, a chat window drawn as an isometric panel with lines of messages on its face, with a callout: Gradio chat, smolagents' built-in chat window, shows thoughts, code and tool calls. Two arrows link it to a large shaded isometric cube in the middle: your question plus listing URL goes to the cube, and the answer, every step shown, comes back. The cube is labelled Manager, a CodeAgent with no prompt of its own, and a callout from it reads human_intervention: asks you on the server's terminal. From the manager, three lines labelled agent(task=…) fan out to three smaller cubes on the right, the specialists, each with its own tools: realtor_agent reads one listing page and works out dollars per square foot; comparable_agent searches the web for similar homes, with links; document_agent searches 14 contract forms and says which apply. A return arrow runs from the specialists back into the manager, labelled report, comps, which forms. A title block at the top reads Realtor Agent. Built: 22 Feb 2025, in one day. Agents: smolagents CodeAgents. Model: GPT-4o via LiteLLM, shared. Tracing: OpenTelemetry to Arize.
    • smolagents
    • Python
  5. P-05Cloud CFD simulation

    Caspen

    A startup building computational fluid dynamics in the cloud, on OpenFOAM and other tools, so engineers and students can run real-world simulations, from aerodynamics to industrial automation. The demo simulates the flow in a mixer vessel.

    • CFD
    • OpenFOAM
    • Startup

Earlier work

Before agents, I learned deep learning the hands-on way. I built networks that dream up new faces, write TV scripts and translate English to French, and wrote my own versions of research models like Pointer Networks and the Neural Random-Access Machine. Once, at a university hackathon, I built a classroom in virtual reality.