Background
Volkswagen needed a platform where data scientists could analyze autonomous driving data and train ML models. The existing process was manual, slow, and not scalable. At the same time, there was no central management system for virtual test vehicles (digital twins) that would represent sensor and vehicle data in the cloud. FNTIO was engaged as a platform engineering partner and supported the project over 2.5 years.
Challenge
- Data scientists waited weeks for infrastructure
- Manual setup for every new ML environment
- No central management of virtual test vehicles (digital twins)
- Security and compliance requirements of the automotive industry
- Complex permissions model: Different teams and external partners need granular access to vehicle data
- ML workloads with highly variable resource requirements (burst training vs. idle)
Solution
Phase 1 – Virtual Test Vehicles (Digital Twins): Web-based application for virtual test vehicles. Each physical test vehicle gets a digital twin in the cloud that centrally represents sensor data, vehicle configurations, and test parameters. Self-service portal: Teams create and manage digital twins in minutes instead of weeks. Automatic synchronization with physical vehicle data via S3 data lakes.
Phase 2 – ML Environments: Infrastructure provisioning with Jupyter Notebooks, Snowpark integration, and direct S3 access to vehicle data. Scaling to 100+ parallel ML environments with automatic resource management. ML scaling patterns: Burst provisioning for training jobs (GPU clusters on demand), automatic downscaling upon completion. Apache Spark for distributed processing of large sensor datasets.
Permissions Model: Granular access concept based on teams, projects, and data classifications. Each ML environment only has access to the vehicle data authorized for the respective use case. Integration with Volkswagen's identity management for single sign-on and automatic role assignment.
Snowflake Integration: Snowpark as the analytical layer on top of the vehicle data.
Data scientists work directly in Snowflake with Python and SQL, without having to copy data. Secure Views for controlled data sharing between teams.
Why FNTIO?
Volkswagen needed a partner that combines Snowflake expertise with AWS architecture. FNTIO not only delivered the initial platform but stayed on for 2.5 years as an ownership partner. That demonstrates: We don't leave projects after go-live.
Results
| Metric | Value |
|---|---|
| ML Environments | 100+ parallel |
| Use Case | Autonomous Driving Data |
| Setup Time | Minutes instead of weeks |
| Self-Service | Fully automated |
| Digital Twins | Central management of all virtual test vehicles |
| Partnership | 2.5 years of continuous development |
| Burst Training | GPU clusters on demand, automatic downscaling |
Differentiation
- Not just infrastructure: Ownership partnership across the entire lifecycle
- Combination of Snowflake data engineering and AWS platform engineering
- Digital twin concept taken from idea to scalable platform
- Permissions model that meets automotive compliance requirements
- 2.5 years of partnership instead of one-off project work
Your Starting Point: ML Platform with Self-Service
Your data scientists wait weeks for infrastructure and manual setups are blocking your ML projects? Here is how we solve that.
Week 1: Identify bottlenecks. We analyze your current ML workflow: How long does it take for a data scientist to get a new environment? Where are the manual steps? What is blocking scaling? Directly in the system, not in interviews.
Week 2: First self-service environment.
An ML environment that can be provisioned via self-service in minutes instead of weeks. With notebook access, data connectivity,
and automatic resource management. Your team tests with real workloads.
Week 4: Platform foundation productive.
The core architecture for self-service ML environments is in place: Provisioning automated, permissions model implemented,
burst scaling for training jobs set up. From here, you scale to 10, 50, or 100+ environments.
What Volkswagen achieved after this point: 100+ parallel ML environments, setup in minutes instead of weeks, digital twins for autonomous driving, 2.5 years of partnership.
