From self-driving cars to AI avatars: the same problem in a new suit
Perception, trust, and teaching machines to understand humans — why my job barely changed when I left autonomous vehicles for voice AI.
DRAFT — replace with Matthias’s own words.
When I tell people I went from self-driving cars to conversational AI avatars, they hear a career change. I hear a change of costume. Underneath, it’s the same problem I’ve been working on for a decade: building a machine that understands a messy human world well enough to act inside it — and knowing exactly how much to trust what it thinks it understands.
Perception was never really about cars
A self-driving stack spends most of its energy on one question: what is actually out there? Cameras, lidar, radar — each one a partial, noisy witness. The hard part isn’t any single sensor. It’s fusing several unreliable stories into one estimate you’d bet a life on, and attaching an honest confidence to it.
A voice avatar of a real person is the same shape. The “sensors” are now words, tone, and context. The “world” is what the person actually thinks and would actually say. And the fusion problem — retrieval, grounding, synthesis — is still the art of turning partial evidence into a response you can stand behind.
Both jobs come down to one sentence: don’t let the machine be confidently wrong about a human being.
Trust is the product
At Cruise, a false positive and a false negative were not symmetric. Phantom-braking for a shadow and rolling through a real pedestrian are different kinds of wrong, and you tune the whole system around that asymmetry.
At Kugelblitz, the asymmetry is different but just as sharp. We build avatars of real creators, and our guiding principles are quality, truthfulness, and experience — the most truthful virtual avatars of creators. An avatar that makes up a plausible opinion the person never held is our version of the confident wrong answer. So the interesting engineering isn’t making it talk. It’s making it decline to talk when the grounding isn’t there.
The stack looks modern — voice synthesis, RAG, personalization — but the values are inherited straight from perception:
- Represent uncertainty honestly instead of papering over it.
- Prefer a graceful “I don’t know” to a fluent fabrication.
- Measure the failures that matter, not the ones that are easy to measure.
Why the throughline matters
I could tell my story as a series of pivots — AV perception, then industrial vision, then voice AI. But that framing misses the point. I didn’t restart three times. I kept solving one problem and let the domain move underneath me.
That’s a comforting thing to notice, and not only about myself. If your work has a real spine — a question you keep returning to — then changing industries isn’t starting over. It’s carrying the same expertise into a room where fewer people have it.
The suit changed. The problem is exactly where I left it.