Machine Learning · SageMaker foundation platform

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

Agentic workflowsModel evaluationServerless analytics on S3AI-native deliveryDay-2 runbooks & team enablement

400,000

vehicles forecast across — a line of business that did not exist before

75B

data points a day in production

~73%

prediction accuracy (R² 0.73) against a >70% contractual bar

Amazon SageMakerAmazon S3Amazon AthenaAWS Step Functions

Sibros’s AI roadmap runs through Amazon SageMaker. VeUP built the production platform that roadmap stands on: SageMaker training and managed inference over an Amazon S3 / Apache Iceberg / Athena telemetry lake, orchestrated with Step Functions and EventBridge, governed with IAM and KMS, and proven against contractual accuracy criteria at 75-billion-points/day scale — then handed to the customer’s own engineers to run.

The challenge

Every platform company now carries a generative-AI ambition; few carry the production AI substrate it requires. For Sibros — a connected-vehicle OTA and telemetry platform ingesting on the order of 75 billion data points a day — the gap was concrete: telemetry economics needed a deployed predictive model with a contractual accuracy bar, and the longer-term AI roadmap needed what that first workload would force into existence — a governed data lake, a repeatable training path, managed inference with real observability, and a team able to operate all of it without a partner in the loop.

The solution

VeUP delivered the platform end to end on Amazon SageMaker. Training runs over a telemetry lake built on Amazon S3 with Apache Iceberg table format and Amazon Athena query access; AWS Step Functions and Amazon EventBridge orchestrate the pipeline; the production model serves through a SageMaker endpoint behind a REST API with IAM authentication and KMS encryption; Amazon CloudWatch carries the operational telemetry. Model choice honored the customer’s frugality constraint — deliberately simple algorithms (XGBoost, Random Forest, regression) that hit the bar without exotic compute. The first statement of work passed its contractual acceptance gate; a second extended the platform; and a structured knowledge transfer moved independent operation to Sibros’s own engineering team. The result is the SageMaker estate — data, training, deployment, governance, and operating skill — on which the platform’s generative workloads land next.

Production outcomes

KPI	Result
Acceptance	~73% production accuracy (R² 0.73) against a >70% contractual bar on the 75-billion-points/day telemetry workload.
Model quality	Accuracy measured on held-out validation excluded from training, with multi-seed runs to bound seed variance; prior in-house approaches had fallen short of the same bar.
Platform continuity	Multi-SOW engagement with the customer’s team operating the SageMaker pipeline independently since knowledge transfer — the foundation in place for the platform’s generative and decisioning workloads.
Lessons & continuation	GenAI readiness is earned in production: a governed lake, a measured training loop, and managed inference under IAM/KMS are the prerequisites; deliberately simple models that ship beat sophisticated models that don’t.

AWS services in production

Amazon SageMaker (training + endpoint)Amazon S3 (Apache Iceberg)Amazon AthenaAWS GlueAWS Step FunctionsAmazon EventBridgeAWS IAMAWS KMSAmazon CloudWatch