Sign in

Model FactoryTurn your own cases into models that keep improving.

See the whole platform
Example data
Model Factory · prior-auth-specialistTraining
  1. TrainStep 40 of 50
  2. Playground4 checkpoints
  3. DeployStable name
0.20.40.60.8base model 0.39
  • Held-out score0.61
  • Gain over base+0.22
  • TeacherCura 1T

Your cases

The work your agents already do is the evidence a better model needs.

  • Task MapGroups production runs into the kinds of case your agents actually see.
  • Benchmarks from real workDrafts a benchmark from those groups, so a model is scored on your cases.
  • SynthesisGenerates more test cases for an agent where real ones are thin.
Example data
Benchmarks · Task Map563 runs grouped
  • Referral missing the physician's order214 runs
  • Out-of-network imaging requests131 runs
  • Retro-authorization after admission96 runs
  • Duplicate submissions52 runs
Benchmark drafted from 3 clusters120 tasks drawn from real runs

Train and compare

Cura 1T teaches a model specialized on your cases, and every checkpoint is scored against the one before it.

  • Cura 1T as teacherYour cases train your model. They never train Cura.
  • Held-out scoringCheckpoints are scored on tasks the model never trained on.
  • PlaygroundPut a checkpoint and the base model side by side on the same case.
Example data
Model Factory · PlaygroundCase 118

Referral for lumbar MRI, no conservative therapy documented. Recommend?

Base model

Approve: imaging is medically necessary for back pain.

Score 0.20
Checkpoint · step 40

Pend: criteria need six weeks of conservative therapy first. Request the PT notes.

Score 0.90
Serve as prior-auth-specialist Deploy

Deploy what wins

Ship the version that scores best on your benchmark, under a stable name your agents call.

  • ExperimentsScore model and agent versions on the same tasks, reward 0 to 1.
  • DeployServe the model you choose under a stable name your agents call.
  • Inside your boundaryTraining and serving stay within your compliance boundary.
Example data
Benchmarks · Experiment120 tasks

Task Map benchmark · Prior authorization

  • prior-auth-specialist · step 400.84
  • Prior auth intake v12 · general model0.71
  • Base model0.58

Reward from 0 to 1, scored on held-out tasks.

Your models, owned by you

The specialized models built from your cases are yours: trained on your evidence, inside your compliance boundary.

Read more

Build your first model with us

Bring one workflow your agents run every day. We'll show you what a model trained on it can do.