Modern motorsport telemetry systems generate massive amounts of data including GPS traces, IMU measurements, CAN signals, and vehicle dynamics. In most cases, analysis happens after the session, often in the cloud. By the time insights are available, the opportunity to correct driving behavior in real time is already gone. For deterministic feedback during a session, cloud-dependent approaches are too slow, too fragile, and sometimes simply unavailable.

In this talk, we walk through the engineering journey of building a real-time telemetry analysis system that runs entirely at the edge on embedded Linux. The objective was straightforward: detect driving patterns and performance anomalies during a session without relying on connectivity. Achieving that goal required solving a set of practical system-level challenges that extend far beyond data acquisition and model training.

We begin with the development pipeline: training a model offline, exporting to ONNX or TFLite, quantizing for constrained hardware, and deploying to embedded System-on-Modules. We compare CPU-only execution against NPU acceleration, highlighting latency, memory footprint, and sustained-load behavior. Real benchmark results demonstrate where hardware acceleration delivers measurable gains and where it introduces additional constraints.

Running inference once is not the hard part. Shipping a complete embedded systems product is.

The talk then focuses on the integration and production aspects of edge AI systems. We examine kernel driver and user-space runtime alignment, accelerator operator support limitations, memory pressure under sustained workloads, and thermal behavior during continuous inference. We discuss containerized deployment on embedded Linux, using Torizon OS as a reference implementation, including hardware access from containers, separation of sensor ingestion and inference pipelines, reproducible builds, and safe over-the-air model updates without reflashing the device.

By the end of the session, attendees will have a practical blueprint for taking an AI model from experimentation to a production-ready embedded deployment. More importantly, they will gain an honest understanding of what breaks, what scales, and what must be designed early when building real-time intelligence on embedded Linux systems.

This is not a showcase of AI capabilities, but a systems engineering story about building, benchmarking, integrating, and maintaining edge AI under real-world constraints.