What I’m building

Agent-native instruments

Two open-source tools built for my imaging work, released July 2026. Each gives an AI agent a bounded set of verbs over MCP while the human keeps the decisions. The code is public: read it, run it, check the claims.

• CBCTScope

  A complete, local-first CBCT and 2D radiograph viewer, built by the radiologist who reads with it. The first CBCT reading viewer with native AI-agent control: the agent moves the camera, the human reads.

  About the viewer →

• Quoin

  A macOS text editor where AI agents are first-class users. The agent edits, you decide.

  About the editor →

Both live at github.com/rezamotaghi.

Grading AI reports against a two-radiologist reference

My main research work, in public as Radigrade: a registered benchmark for grading AI-written radiograph reports. Each report is compared with a reference read by two oral and maxillofacial radiologists reading together, one axis at a time, with no blended score and no rank. General models are graded on a sealed exam under the registered protocol, and the results, the protocol and the change log are at radigrade.com. The professional record is on the research page.

B2B AI automation

Automation for business workflows, built and shipped in my role as Chief Innovation Officer. My products, the company’s channel, six figures in production: not a portfolio piece. The unglamorous kind of AI: repeatable hours now run unattended, and the human sign-off stays exactly where it earns its keep. Most of what the essays claim about reliability was earned here first.

Essays on imaging AI, evaluation, and building alone

A few pieces a month, published here and sent to subscribers first. No framework gets a name before it has survived real work.

What I’m not doing

• Chasing model releases. Tools earn their place slowly here; the 2-Tool Rule is house policy, not just an essay.
• Hype. If a claim would not survive a skeptical radiologist, it does not ship.
• Scaling for its own sake. One person, deep focus, on purpose.