Tyriad

I co-founded Tyriad in 2021 with two friends I grew up with. Most automotive LiDAR at the time was expensive, mechanical, and fragile. Our bet was a sensor that could be cheaper, more durable, and see more, paired with the software that makes the data useful.

Time-of-flight LiDAR fires a pulse and times the echo. FMCW sweeps the laser's frequency and mixes the return with the outgoing light. The beat frequency encodes range, and the Doppler shift gives the velocity of every point from a single measurement. It's also immune to crosstalk from sunlight and other sensors, and holds up better in bad weather.

Traditional LiDAR stacks many lasers in a housing that spins, and moving mass is what breaks first on a car. Our design, Pulse Mk 1, used a single laser and moved only a small dual-axis scanning mirror, making it closer to solid state and much less prone to wearing out.

The wavelength mattered. Light at 1550nm is absorbed before it reaches the retina, so IEC 60825 eye-safety limits allow far more power than at the 905nm used by many pulsed systems. More power means more range while staying eye-safe.

A point cloud is only useful if the car understands it. We trained semantic classification models that went past labeling objects to reasoning about what might happen next. A basketball rolling into the road isn't a hazard by itself, but a child might follow it, so the car should slow down. Switch the point cloud above to Semantic to watch it play out.

Today that sounds like something you'd ask an LLM. We were doing it in 2021, with purpose-built models and our own data, alongside real-time data processing on AWS Lambda and IoT Core.