Why should you care?

AIPredict trains fraud models on a GPU worker while many API pods only load zips from blob storage.

See it live — copy this example

Use a .NET console or Web API project with Microsoft.ML. Run dotnet run after pasting.

// Worker writes blob; API pods read same URI
var blobUri = Environment.GetEnvironmentVariable("MODEL_BLOB_URI");
await using var stream = await blobClient.OpenReadAsync();
var model = ml.Model.Load(stream, out var schema);
services.AddPredictionEnginePool<TxRow, FraudPred>()
    .FromFile("/models/cache/fraud.zip");

What happened?

• Train centrally, publish artifact to shared storage, scale stateless inferencers horizontally.
• Follow the steps below — typing the code yourself is the fastest way to learn.

Practice next

1. Worker trains and uploads zip.
2. API downloads to local cache.
3. Register pool from cache path.
4. Add ETag check for reload.
5. Use regional blob replicas.

Remember

Central train. Shared artifact. Stateless infer pods.

AIPredict distributed infer

Three API regions load same fraud zip.

Outcome: Consistent scores without per-pod training.