A model per stage
Each clear tray moves the teeth a small step, and each step needs its own arch model.
Change the plan in a sentence.
Every stage follows.
Every plan change
means rebuilding the CAD.
After diagnosis, dentists still turn the plan into per-stage 3D models by hand. A treatment longer than a year needs dozens of them.
Dozens per case
Longer treatment means more stages, and more CAD work.
Rework on every change
A revised plan sends you back into the existing design.
From the field
Site visits and consultations with a board-certified specialist in integrated dentistry confirmed this repetitive CAD burden.
Ask for a change.
Get a new, checked plan.
Each step shows who acts. The agent proposes, and every call that matters stays with you.
- 01
Describe
You set the prescription in plain words.
You decide - 02
Draft & check
The agent drafts every stage, and geometry tools flag movement, collision and space problems.
Agent proposes - 03
Revise
You ask for changes. Conditions you already gave carry over.
You decide - 04
Approve & export
You approve a plan that passes the rules. Print files come after that.
You decide
-
Dentist
No extraction, keep the canines fixed.
-
Agent
Got it. I set the target arrangement and drafted every stage with the canines held in place.
-
Agent
Two teeth collide in the middle stages, marked in red. Allowing IPR would make room while the canines stay fixed.
-
Dentist
Allow IPR. Keep everything else.
-
Agent
New plan. Kept No extraction Canines fixed Added IPR
It passes the rule check. Review it and approve when you're ready.
A scripted replay. The 3D views are rendered from plans cuAlign computed for two public sample scans. The conversation was written for this page.
See every stage
before you sign off.
Drag through a real 18-stage plan. Red marks a collision the rule check found.
Nothing leaves
without your approval.
Diagnosis, prescription and the final choice stay with the dentist.
Drafts only
Every plan is a proposal until you approve it.
Your conditions hold
The agent never loosens them to pass a rule.
Explicit approval
Plans with violations can't be approved or exported.
No patient identifiers
Patient registration rejects identifying details. Chat messages with ID or phone numbers are refused.
No certification yet.
The tests are public. Re-run them.
Everything below is in the public repository and can be re-run.
Public scans
Sample cases are real upper-arch scans from the open Poseidon3D dataset (CC BY 4.0).
Geometry checks
Plans are checked for space shortage, collisions, movement per aligner and stage count. A benchmark runs those checks on five synthetic presets.
19 scripted scenarios
They cover revisions, plans that honestly fail, unsupported requests, and diagnosis questions the agent hands back to the dentist.
One dentist, 12 cases
A dentist judged space and approach for 12 sample scans by eye. We treat it as a qualitative reference and report no score.
The automated tests (uv run pytest) and a rule-based plan (cualign plan) run offline. The verification record in the source repository lists what passed and what hasn't been checked. Clinical validity is not evaluated.
Works from per-tooth STL,
hands back STL for every stage.
Input
Segmented upper-arch scan
One STL per tooth, numbered in the Universal system, plus an optional gum scan.
Output
ZIP of stage files
Per-tooth STL for every stage and, with a gum scan, a printable arch model per stage.
Not yet tested with specific intraoral scanners, CAD software or 3D printers. Automatic segmentation of one-piece scans is not supported.
The NVIDIA stack underneath,
and what each piece handles.
Nemotron via NIM
Understands your prescription and chooses the tools.
NeMo Agent Toolkit
Runs the planning agent and its geometry checks.
NeMo Guardrails
Checks the agent's input and output on each call.
Agent Skills
The planner reads a clinical-rules skill before it plans.
Sandboxing with OpenShell and MCP access through NemoClaw are experiments.
Who builds cuAlign.
Hackathon team
The cuAlign team, building for the NVIDIA Hackathon 2026.
Clinical input
Consultations with a board-certified specialist in integrated dentistry, and a dentist who reviewed the sample scans and wrote their prescriptions.
What it handles today,
and what it doesn't yet.
- Upper arch only. Translation, incisor rotation and vertical correction.
- Not yet modelled. Rotation of canines and back teeth, torque, the lower arch, occlusion.
- Computed checks. IPR adjusts space only in the calculation, and the collision check doesn't guarantee zero contact for every tooth pair.
- Hosted or local. A hosted prototype is available on request. You can also run it on your own machine.
- For review. Exports are not validated for manufacturing or wear.
Questions.
Is cuAlign a medical device?
No. It is a research prototype and must not be used for diagnosis or treatment decisions.
Which cases does it not support?
Lower arches, rotation of canines and back teeth, torque, occlusion, and one-piece scans that are not segmented per tooth.
What happens to scan data?
On the hosted prototype, scans and plans are stored on that server, behind access credentials. Run it locally and they stay on your machine. The agent conversation goes through the NVIDIA API either way, and there is no per-user isolation, so don't upload identifiable patient data.
How much does it cost?
Nothing. cuAlign is a research prototype and isn't sold.
What are the terms?
Use is covered by the End User License Agreement: research and demonstration only, no warranty. Sample scans are CC BY 4.0, and third-party components keep their own terms.
What do I need to run it?
Nothing, if you use the hosted prototype. We share its access details on request. To run it yourself you need Python 3.12 and uv. Plans compute without a key, and the conversation needs an NVIDIA API key. The Quickstart walks through it. The current prototype interface is in Korean.
Try a plan change
on a sample case.
We share access to the hosted prototype on request. The docs cover running it yourself.