Argon designs and fixes. My checker decides. You sign off.
The loop: Argon watches the operator, writes the holder, looks at the collision and rewrites the design, up to three rounds, until my checker passes it or the time cap ends it, then stops for the engineer.
Explained simply
Building with LEGO, you make a little stand to hold one piece still while your other hand uses a screwdriver. If the stand blocks the screwdriver, you have to take it apart and rebuild it. FixtureFit finds that problem on the computer, before anyone 3D-prints the stand.
The object the worker is assembling. It must not move.
A 3D-printed holder shaped to keep the part still.
The tool that tightens the bolt. It needs a clear path to reach it.
Five steps
- Show the job: the part, the tool, and a short clip of the operator using it.
- Argon designs the holder: it writes the holder as CAD code.
- The checker tests the wrench's path: it moves the modelled wrench along its path and measures the gap to the holder.
- Argon fixes collisions: if the wrench hits, Argon changes the design and the checker tests it again.
- The engineer approves: only then can the holder be ordered as a test print.
What the numbers mean
−1.03 mm: the wrench and the holder overlap by 1.03 mm in the model. A collision.
+1.97 mm: a 1.97 mm gap, more than the 1.5 mm required. Clear.
A PASS is about the modelled tool path only: not print tolerances, not how strong the holder is. That's why an engineer signs off.
Who does what
Gemini 4 Argon: the AI engineer. It reads the clip, designs the holder and fixes it.
Gemini Managed Agents: Argon's workshop in Google's cloud, where each fix runs.
The checker: the referee. FixtureFit's own code; it alone decides PASS or FAIL.
Gemini 3.8 Flash: writes one plain sentence about a failure. Never the verdict.
Gemini Omni: made the sample operator clip (labelled AI-generated).
Required margin 1.5 mm. Negative = the wrench hits the holder by that much. The repair was measured live: 3 of 3 passed in 70–87 s.
Argon's four jobs
Watches the operator clip, writes the keep-clear rules
Argon reads the video and returns each moment the tool engages: the time, the tool, the direction it moves, and the face it reaches, plus what must stay clear. It never states a dimension: every millimetre comes from the CAD. The engineer confirms each rule before it is used.
Writes Draft 1 as CAD code, in one pass
From the confirmed rules, the part and the tool, Argon writes the whole fixture program (a cadquery file with its parameters). The app builds it and my checker judges it exactly as written.
Looks at the collision and rewrites the design
When a change makes the wrench hit, Argon gets the failing number and a picture looking down the tool's path. One saved agent runs twice at the same time with two strategies (move a wall / cut a relief). Each run is a fresh copy of the agent's environment with its network switched off. Each sends back new CAD code; the app builds and checks both and keeps the one that passes with the larger gap.
Stops for the engineer's sign-off
Each fix ends on a sign-off request: the agent pauses. Only a fixture my checker passed can be approved; the approval is recorded and answers the agent's pause. Then the test order.
The helpers, and the judge
One-line summary
Says in one sentence what the failing check means for the operator. Labelled “companion summary — not part of the verdict”; if Flash is slow, the checker's own sentence shows instead.
Made the sample clip
The sample operator clip is AI-generated from my render of the part (first frame), labelled “AI-generated clip (Gemini Omni)”. You can upload your own clip during onboarding.
Decides, outside Argon's sandbox
Re-imports every fixture with its own tessellation and checks it on its own copy of the part, tool and spec: tool clearance along the path, no overlap with the part, locating points touching, camera view clear, one closed solid, fits the printer. Any code Argon writes is checked for forbidden imports before it runs.
What is live and what was made before judging
| Step | On screen |
|---|---|
| The checker, every number | Live, in the app |
| Argon's fixes (two forks) | Live with a key, green LIVE badge and the interaction id. Without a key or past a limit: a recording labelled REFERENCE RUN |
| Draft 1 of the sample job | Made before judging, labelled with its interaction id and seconds. Onboarding can write a new one live (~2 min) |
| Reading the sample clip | Read before judging, labelled. Your own clip is read live |
| Stripe | Test mode only. No vendor connected; nothing is printed |
Measured honestly: in a pre-registered baseline every approach passed every job (the test hit the ceiling), so I don't claim the loop or the picture makes Argon better. The claim is the independent check. See Evals.