Machine Learning Applications · Connected-Vehicle ML

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

74%

prediction accuracy against a >70% contractual target

Amazon SageMakerS3 + IcebergAWS GlueAmazon AthenaStep Functions

Production predictive ML platform on Amazon SageMaker for connected-vehicle telemetry at 75 billion data points per day — enabling a new AI/ML line of business for Sibros's international-automaker customers.

The customer

Sibros is a North-America-headquartered ISV operating in the connected-vehicle and over-the-air (OTA) software-update space. Their unified OTA-updater and telemetry platform sells to automotive OEMs and fleet operators worldwide, ingesting ~75 billion telemetry data points per day from fleets ranging from 20,000 to 400,000 vehicles each.

The challenge

Sibros's commercial model and AWS infrastructure cost are both directly driven by how much telemetry each vehicle logs — which is in turn driven by customer-defined logging rules. A given fleet configuration can produce anywhere from ~13 billion to ~40 billion data points per day. Before engaging VeUP, Sibros had no model-driven way to forecast fleet data volume before rollout, creating mispriced rollouts, AWS-cost variance, and customer trust risk on cost estimates.

Sibros's engineers had already tried to crack it themselves — XGBoost and Random Forest models running on laptops, then GCP BigQuery ML. Every attempt stalled below the 70% accuracy the business needed before it could put a forecast in front of a customer.

The solution

VeUP built the training and inference platform inside Sibros's own AWS account: Amazon SageMaker for managed training, the model registry, and real-time inference, sitting on an Apache Iceberg data lake in S3 with AWS Glue and Amazon Athena as the feature layer, orchestrated by Step Functions and EventBridge, secured with IAM Identity Center and KMS, watched by CloudWatch.

The decision that made the accuracy stick: strict GroupKFold cross-validation on rule_id across 30 seeds, so the model has to generalize to logging rules it has never seen — exactly the situation it faces every time a customer adds a new one.

The platform shipped at 74% production accuracy on the 75-billion data-points-per-day workload, clearing the 70% acceptance bar the engagement was contracted against. A second modelling pass under the same cross-validation discipline lifted production R² from 54% to over 80%. Sibros's engineers now run the pipeline themselves.

Architecture

The earlier attempts — local Python ML and GCP BigQuery ML — both capped below 70% accuracy at production scale, so VeUP built new on AWS: SageMaker training and real-time inference over the Iceberg-on-S3 data lake, with the Amazon Bedrock AgentCore OTA-decisioning pipeline calling the SageMaker endpoint as its prediction tool.

The architecture on AWS — SageMaker training and inference over the Iceberg-on-S3 data lake, with the AgentCore pipeline calling the endpoint as its prediction tool.

Where it started

Assessed baseline · new implementationConnected-vehicle telemetry — ~75B points/day · Fleets of 20K–400K vehicles

Starting point

Prior attempt: laptop-scale Python ML

scikit-learn, XGBoost, and Random Forest models with ad hoc feature engineering on engineer laptops — not production capacity.

Starting point

Prior attempt: BigQuery ML

Regression models evaluated on GCP BigQuery ML as an alternative platform — also at non-production capacity.

Gap

Both capped below the 70% accuracy bar

Neither prior attempt could generalize to ~75B telemetry points/day of forecasting across 20K–400K-vehicle fleets.

Constraint

Three weeks, managed services only

A 3-week delivery clock and a frugality constraint — managed services only, no exotic infrastructure.

Driver

Price OTA rollouts from logged data

Forecast logged-data volume by vehicle type, logging config, and time window to price per-fleet OTA rollouts.

A new implementation on AWS — "a concept move," not a lift-and-shift; both prior systems were retired.

Production outcomes — customer-business KPIs

KPI	Result
Production prediction accuracy on Amazon SageMaker	74% production accuracy on the 75B data-points/day workload, against a 70% acceptance bar. R² 0.8211 on the held-out test set; 88.9% of predictions within 2× actuals; leave-rules-out R² 0.73; overfitting gap −0.06.
New AI/ML telemetry line of business	Production SageMaker platform enables Sibros to offer predictive AI/ML telemetry products to international automaker OEMs as a new line of business that did not exist before this engagement. A v2 modelling effort lifted production R² from 54% → 80%+, expanding the LOB's accuracy headroom.
Customer-team independent operation	Sibros's engineering team runs the production Amazon SageMaker pipeline on its own — knowledge transfer was part of the delivery, not an afterthought.

Where it went next

After accepting the first model in production at 74% accuracy, Sibros came back for more: a second engagement extended the platform in spring 2026, and the v2 modelling work pushed production R² past 80%.

AWS services in production

Amazon SageMaker (training + real-time endpoint)Amazon S3AWS Glue Data CatalogAmazon AthenaAWS Step FunctionsAmazon EventBridgeAWS IAM Identity CenterAWS KMSAmazon CloudWatchAWS Cost Explorer

Integrations

Apache Iceberg