Skip to main content
This walkthrough demonstrates the modeling assistant skill from start to finish. Starting from an empty project with a single dataset, the coding agent analyzes the data, generates training scripts, trains multiple models as Domino Jobs, and compares results in Experiment Manager.
This walkthrough follows the AI Coding Assistants quick intro video. You can watch it alongside this guide for the full experience.

​
Setup

This demo uses a Domino project with a single dataset for predicting diabetes. The project has no existing code or models.
  1. Create a new Domino project or use an existing one with a dataset.
  2. Launch a VS Code workspace using the Domino Standard Environment.
  3. Open Claude Code and authenticate (as explained in Get started).

​
Step 1: Activate the modeling assistant

In the Claude Code prompt, type: 
Use the modeling assistant skill. Do a quick analysis of the data set
in the project and then train a few models and validate.
  • Claude Code activates the modeling-assistant skill and begins by detecting the Domino environment.
  • The agent identifies the project, locates the dataset, and starts the workflow.

​
Step 2: Train models as Domino Jobs

The agent writes a data analysis script, generates a parameterized PyTorch model, creates a training script with MLflow tracking, and submits training as Domino Jobs. All steps run automatically. The modeling assistant creates model architectures, submits training as Domino Jobs, and monitors progress Key details about the generated training pipeline:
  • Training runs as a reproducible Domino Job, not inline in the workspace.
  • All parameters, including batch size, learning rate, epochs, and architecture, are logged to Experiment Manager.
  • The agent trains multiple models with different configurations for comparison.

​
Step 3: Review results in Experiment Manager

Experiment Manager tracks each training job with complete parameter lineage. Navigate to Experiments in your project sidebar to see the runs. The experiment view displays:
  • Training loss and accuracy curves over time.
  • Model parameters, including activation, batch size, dropout, epochs, and hidden layers.
  • Full lineage back to the training job and code.
Experiment Manager shows training loss, accuracy curves, and all parameters for each model run

​
Step 4: Compare models

After all models finish training, the agent generates a comparison grid showing both models side by side with key takeaways: The modeling assistant compares both models and provides recommendations based on accuracy, efficiency, and parameter count In this example, the agent trained two models:
Metricsmall_32x16 (865 params)medium_64x32x16 (3,297 params)
Test Accuracy96.19%96.46%
Test Precision99.81%99.97%
Test Recall94.92%95.15%
Test F197.30%97.50%
The agent concluded the small model is the better production choice given the marginal accuracy difference and significantly fewer parameters.

​
Step 5: Register a model

From the Experiments view, register any trained model to the Domino Model Registry by clicking the Register button on the experiment run.
All training jobs reproduce consistently. Re-run any job from the Jobs page to verify or iterate on results.

​
Summary

In this walkthrough, the modeling assistant skill handled the complete workflow:
  1. Detected the Domino environment and project structure.
  2. Explored and analyzed the dataset.
  3. Generated a parameterized PyTorch model and training script.
  4. Submitted training as reproducible Domino Jobs.
  5. Tracked all parameters and metrics in Experiment Manager.
  6. Compared multiple model configurations with recommendations.
All lineage stays intact. The Domino platform tracks, reproduces, and audits every step from data analysis to model comparison.

​
Next steps