Without a good foundation, everything erodes

As you can imagine, there is a lot of math behind building a coaching engine which uses car telemetry data to give you insights about how to improve your driving. Every sensor read is used in a multitude of places, and slight errors in how the sensor data is used can create undesirable results.

I've been spending some time re-tooling some of the math and cleaning up how the data is filtered to make sure that the foundational layer of the coaching engine is as accurate as possible. It gets tough when each car has slightly different sensor inputs, and something simple like an alignment change can create a whole host of issues. This was one of my latest battles. The tool now calibrates the steering angle based on the readings where the steering wheel is near center and there is no yaw rate in the car. I've seen cars with 10-degree steering offsets just to keep it pointed in a straight line.

It also exposed some other areas where there were opportunities to fit a better model to the tires and use more accurate constants. A new release with this tuning should follow soon.

The next big chunk of work is improving the wording of the automatic insights as well as making them deeper and richer. Version 1 of the insights was corner-focused. I've since added the ability to consider long straights, so now you'll get coaching on corners where exit speed matters most, because of what follows them.