Sibros: a 75-billion-point/day SageMaker model + Bedrock AgentCore pipeline

In a connected-vehicle platform, the logging rule is the throttle on everything downstream: each customer-defined rule determines how much telemetry every vehicle emits, and therefore what the platform’s cloud bill — and its customers’ bills — will be. At 75 billion points a day across fleets of hundreds of thousands of vehicles, misjudging a rule’s data volume is an expensive mistake discovered weeks later on an invoice. Sibros needed the IoT loop closed: a way to predict, at configuration time, what a logging rule would cost in data before a single vehicle started emitting under it.

VeUP built a production predictive-analytics pipeline over the platform’s telemetry lake. The fleet’s emitted telemetry lands in Amazon S3 under the Apache Iceberg table format with Amazon Athena as the query layer; Amazon SageMaker trains telemetry-volume-forecasting models over it — deliberately simple algorithms (XGBoost, Random Forest, regression), honoring the customer’s frugality constraint at this data scale — and serves them through a SageMaker endpoint behind a REST API the platform calls at logging-configuration time. Validation discipline matched the fleet structure: strict GroupKFold cross-validation by logging rule, so the model is scored on rules it has never seen, the way production will actually use it. The pipeline was delivered against a contractual accuracy bar, extended in a second engagement that materially improved generalization, and transferred to Sibros’s engineering team, who operate it independently.