Mark Zeidan

I build small, useful things with AI.

Products, agents, and the occasional piece of hardware. Below is what I’ve built, and what I learned building it.

Mark ZeidanNorthern Virginia
A 3D model of Algonkian Golf Course floating above the clouds, with fairways, ponds, and thousands of trees
Algonkian StarterAll 18 holes, built from OpenStreetMap, with a simulated Saturday of 22 groups. Best on a laptop.
A phone showing a lit 3D kitchen with finish options, between panels explaining what the guide is doing
Home ConciergeA home-buying guide that prices every choice monthly and walks you through a lit 3D house. All homes are fictional.
Apex Arena: 1776An Unreal Engine shooter where you reload by answering a dispatch from the General. Built in Unreal Engine 5.

Work

12 entries
  1. 01

    Algonkian Starter

    An 18-hole model of a real golf course, built from OpenStreetMap, for keeping a busy Saturday on pace.

    3D simulationPrototype
  2. 02

    Home Concierge

    A mobile home-buying guide that turns every choice into a monthly cost, then walks you through a lit 3D house.

    Interactive prototypePrototype
  3. 03

    Gonkbot

    Log golf rounds in ChatGPT or Claude while you play: scorecards, trends, and handicap math.

    MCP serverLive
  4. 04

    Sky Lab

    A narrated iPad lab that shows young kids how helicopters fly. Blow into the microphone to make wind.

    Learning appLive
  5. 05

    Apex Arena: 1776

    A first-person shooter in snowy 1776 Trenton where every reload is a history question. Miss one and your musket stays empty.

    Game · Unreal EnginePrototype
  6. 06

    Agent Macropad

    A hand-wired 14-key pad with a knob for steering coding agents. Case based on a design by Newtle Kim.

    HardwareIn progress
  7. 07

    Birdcam

    A Raspberry Pi feeder camera. Claude names the species and deletes the frames without a bird.

    HardwareOpen source
  8. 08

    Homework Helper Arm

    A robot arm that hears a question, looks over the table, and points at the answer as it explains.

    RoboticsPrototype

Code and experiments

  1. 09

    Family Steward

    A household-help agent where the model proposes and people approve.

    AI agentOpen source
  2. 10

    Hermes Agent

    Pull requests to Inkbox’s fork of the Hermes agent that make its SMS channel dependable: failures, bursts, follow-ups, long messages.

    Open source6 PRs merged
  3. 11

    Drone lab

    An MCP server that lets an agent fly PX4 drones in simulation: missions, failsafe tests, and telemetry checks.

    Robotics · simulationSimulation
  4. 12

    Local model lab

    Benchmarking and fine-tuning open models on one GPU. Teaching a model to use a lookup tool beat training more facts into it.

    ML experimentsWorkstation GPU

Ideas I keep coming back to

Keep the policy out of the model.

The model proposes. Approvals, retries, and the audit log live in plain code it can’t talk its way around.

Ship tools, not inference.

Gonkbot runs no model at all. The golfer’s own assistant does the thinking; the server keeps the books.

Cheap hardware, smart model.

A Pi, a camera, and a vision model go a long way. Neither the birdcam nor the robot arm trains anything.

Guardrails come from failures.

Every rule my deploy agent follows is there because it broke something first.