Model development and deployment

Select for the task

Forecasting, image recognition, speech processing and generation need different data and evaluation methods. We define the expected output, latency limits and consequences of important errors before comparing suitable approaches.

The work can involve training on company history, adapting an existing model or configuring its use without further training. We review the proposed model’s usage terms and whether the required data is available for the agreed task.

Keep training and evaluation separate

We prepare datasets, review labels and reserve examples for independent evaluation. The split accounts for time, related documents, customers or other connections between records so that it tests performance on new situations.

We compare quality across important groups alongside processing time and resource use. Language models are tested on representative requests; forecasts are checked against held-out periods. If we change model size or format to improve performance, we repeat the quality evaluation.

Prepare for operation and change

The model can run as a scheduled batch process or request-based service in a cloud, private cloud or customer-managed environment. We configure the interface, access, versioning and collection of operational metrics.

After launch, we monitor changes in data and task results as feedback becomes available. Retraining and updates pass an agreed evaluation before release, with a rollback path. For environment sizing, see infrastructure planning.