Agentic AI · Agentic engineering delivery

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 SageMakerAmazon S3AWS GlueAmazon Athena

This is the story of how the Sibros platform got built, and what it sets up next. VeUP ran its agentic-AI-coordinated engineering practice — AI coding agents in MCP-orchestrated build/test loops with Amazon Q Developer as the code-review gate — to ship a production Amazon SageMaker pipeline and REST inference API for connected-vehicle telemetry at 75-billion-points/day scale, inside a single three-week engagement, for a customer that had evaluated multiple competing partners first.

The challenge

Sibros, a connected-vehicle software platform, needed a production predictive-ML capability on a startup timeline — a hard accuracy bar written into the contract, a three-week window to hit it, and a strong preference for a system its own engineers could run without a partner on retainer. Sibros collected bids and evaluated multiple partners before selecting VeUP. The test wasn’t whether a model could be trained; it was whether a partner’s engineering practice could design, secure, deploy, and hand over a production AI system at that velocity.

The solution

VeUP delivered with the practice the Agentic AI competency is about: agentic coding tools running MCP-orchestrated build/test loops as the engineering engine, with Amazon Q Developer reviewing every change — an AI-coordinated loop with human engineers steering scope, architecture, and acceptance. That loop shipped the production system: Amazon SageMaker training and a managed endpoint over an S3/Apache Iceberg telemetry lake cataloged with AWS Glue and queried through Amazon Athena, fronted by a typed, IAM-authenticated REST API and instrumented with Amazon CloudWatch. Delivery didn’t end at go-live: the engagement ran through structured HyperCare and a full knowledge transfer, after which Sibros’s engineering team took over independent operation of the pipeline — including the follow-on generalization work that lifted v2 model quality. And the endpoint itself is deliberately agent-shaped: a typed, authenticated, observable inference tool — the foundation for the OTA-campaign decisioning agent on the customer’s roadmap.

Production outcomes

KPI	Result
Delivery velocity	First partner to ship the predictive-volume model to production — 74% accuracy against a >70% contractual target, inside a single three-week engagement.
Capability transfer	Engagement completed HyperCare and full knowledge transfer; the customer’s engineering team has operated the production pipeline independently since — a transferred operating capability, not a handed-over artifact.
Quality continuation	v2 generalization work under strict GroupKFold cross-validation lifted model R² from 54% to above 80%; a second statement of work signed to extend the platform.
Lessons & continuation	An agentic build loop with a hard review gate compounds speed without sacrificing rigor; shipping the inference layer as a typed, authenticated tool means the ML platform doubles as the tool surface for the customer’s planned OTA-decisioning agent.

AWS services in production

Amazon SageMaker (training + endpoint)Amazon S3 (Apache Iceberg data lake)AWS GlueAmazon AthenaAmazon Q Developer (delivery practice)AWS IAMAmazon CloudWatch