Sibros Technologies Inc.
Using an agentic-coordinated engineering loop, VeUP delivered a production Amazon SageMaker predictive-ML platform for connected-vehicle telemetry at ~75 billion data points per day — hitting 74% accuracy against a >70% target in a single 3-week engagement. Sibros's own team has operated it independently since March 2026.
The customer
Sibros Technologies is a connected-vehicle ISV building an over-the-air (OTA) software-update system and a vehicle-telemetry platform for OEMs and fleet operators. Its platform collects sensor data based on customer-defined logging rules, lands it in an Apache Iceberg data lake on Amazon S3, and exposes it for downstream OEM analytics. Fleet sizes among Sibros's customers range from roughly 20,000 to 400,000 vehicles, with aggregate platform scale on the order of 75 billion telemetry data points per day.
The challenge
Sibros needed to forecast logged-data volume from a logging configuration before a fleet is deployed — turning rollout pricing into something both Sibros and its OEM customers could predict in advance. Prior internal attempts and earlier modeling approaches had stalled short of the required accuracy bar against telemetry at extreme scale. Sibros evaluated multiple competing partners and collected bids before selecting VeUP to build a production-grade predictive model that could clear a contractual accuracy target the team had not been able to reach on its own.
The solution
VeUP delivered the engagement through its agentic-coordinated engineering delivery practice: Cursor and Claude Code driving MCP-orchestrated build and test loops, with Amazon Q Developer for AI-assisted code review on the AWS-native surface. This agentic loop coordinated the AWS SDK calls, Athena queries, Iceberg-table introspection, and SageMaker training and endpoint promotions that took the build to production — with human engineers reviewing every model promotion and IAM-affecting action through human-in-the-loop gates.
The production architecture moves telemetry from the S3 + Apache Iceberg data lake through an AWS Glue and Amazon Athena query layer, into feature-engineering notebooks, then into Amazon SageMaker for training (XGBoost / regression / Random Forest variants) and a hosted SageMaker endpoint fronted by a REST API for on-demand volume prediction. Managed, serverless, and least-privilege-by-default choices were enforced throughout — including strict GroupKFold cross-validation on rule IDs so the model is validated against rules it has never seen at training time. The agentic loop reached the customer's AWS account only through a customer-issued, prefix-scoped IAM role assumed via short-lived STS sessions; no long-lived credentials were ever held.
The resulting SageMaker endpoint is a typed, IAM-authenticated, observable inference tool — exactly the shape a future agent invokes — making it the agent-ready data foundation for Sibros's roadmapped OTA-decisioning agent. Critically, VeUP did not hand over just a model: it transferred a production AI operating capability. Sibros's own engineering team has run the production pipeline end-to-end, independently and without VeUP attendance, since 2026-03-03, with knowledge transfer complete.
Production outcomes
| KPI | Result |
|---|---|
| Prediction accuracy vs. contractual target | 74% achieved against a >70% acceptance bar, in a single 3-week T&M engagement. |
| Generalization R² lift (denser, harder dataset) | 54% → 80%+ R² under strict GroupKFold validation, across 40 → 768 distinct rules. |
| Platform telemetry scale | ~75 billion data points per day. |
| Customer-operated handover | Sibros team running the production pipeline independently since 2026-03-03. |
Customer voice
Sibros selected VeUP after evaluating multiple competing partners and collecting bids, and VeUP was first to ship the predictive-volume model to production at the required accuracy. Sibros engineering leadership has described a roadmap toward systems that can reason about an in-flight OTA software-update campaign — the kind of decisioning agent the production SageMaker endpoint VeUP delivered is built to serve as a foundation for.
Amazon SageMaker (training + hosted endpoint) · Amazon S3 (Apache Iceberg data lake) · AWS Glue Data Catalog · Amazon Athena · AWS IAM (least-privilege roles, STS short-lived sessions) · Amazon CloudWatch.