Qokedas

Qokedas turns raw signals from scientific instruments into structured training data, delivered as reinforcement learning environments. The San Francisco company wants to teach AI models to interpret physical phenomena and operate the instruments that measure them.

Most of the world's data originates as physical signals passing through instruments, and the founders argue this vast pool remains largely untapped by AI still trained overwhelmingly on digital text. As models saturate available online data, measurement offers a fresh, hard-to-fake training signal.

The founding team — Wilhelm Hedenskog, Jesper Johansson and Elias Reinfeldt — started Qokedas in 2026 on the observation that current AI is fixated on digital data while the analog physical world stays out of reach.

The team is building in the open: they've published 'aubench-lite', a benchmark testing an AI agent's ability to decode data from a 1,024-microphone acoustic array from raw signals. They're also part of Y Combinator's Fall 2026 batch.