# Pascal ML

Source: https://www.trypascal.io/platform/ml

Build predictions around real decisions.

Define what a useful prediction must do. Train candidates, compare their results, and test the failure cases that matter before proposing a release.

## Capabilities

### Start with the prediction

Define the target, decision, evaluation metrics, and operating constraints before selecting a model or training approach.

### Prepare data without leaking the answer

Choose governed inputs, define features, and preserve the separation between training and evaluation data.

### Run comparable experiments

Keep datasets, feature choices, parameters, and environments connected to each training run so candidate results can be compared.

### Evaluate against the real objective

Review model quality alongside latency, cost, and the constraints of the business decision. Keep a baseline for comparison.

### Deliver an exact model version

Connect a selected release to batch or online prediction. Retain the evaluation and training lineage behind the deployed model.

### Learn from operating feedback

Connect observations and labels to monitoring. Use performance changes to guide a reviewed decision about evaluation or retraining.

## Connected context

Use Data Infrastructure and Data Pipelines for inputs, Compute for jobs, and Models to connect the objective, evaluation, and release decision. Keep the training data and results available for review.

- Prediction objectives, training data, features, and held-out cases
- Training jobs, candidates, metrics, and evaluation results
- Model versions, prediction outputs, and monitoring evidence

## Controls and permissions

Compare candidates against the same criteria and inspect performance on critical groups. A better overall score can still hide an unacceptable failure. Selecting a candidate, releasing it, and putting it into use are separate decisions.

## Example

Missing a hazardous shipment can matter more than a small gain in average prediction quality. Meridian’s evaluation makes that tradeoff visible: Trial 3 improves the overall score but misses the recall requirement for lithium loads. The team keeps the decision and the reason to continue the study.

## Workflow

### Define the prediction and its use

- Predict which shipments may miss the 21:00 cutoff using information known at the decision time. The intended use is human-reviewed recovery.

### Compare the same evaluation

- Inspect three candidates against the same held-out data. Trial 3 leads overall, but its hazardous-load recall is 84%, below the 90% minimum.

### Record the next decision

- Continue study retains the reason to improve hazardous-load recall. The comparison does not select a new model release.

## Next steps

Define the target, the information available at decision time, and the mistakes that would make a model unsuitable for the work.

- [Contact sales](https://www.trypascal.io/contact)
- [Pascal Work](https://www.trypascal.io/solutions/pascal-work)