Translucent AI

What makes it different

AI agents can now propose hardware designs faster than any lab can test them, and the bottleneck is the test bench. Translucent, built by Yashas Gunderia, does one job: it chooses the next physical test. The agent picks the few tests that settle a design and writes down a forecast for each one before the result comes in.

The team sets cost, time and run-count limits once; solver tests run on their own inside those limits, while bench tests always go to an engineer with a procedure attached. When the data cannot support a design, the agent asks for the measurement instead of guessing — and every step goes into a record anyone can re-check offline.

Batteries are the first system. In a simulated fast-charge race with 8 experiments, Translucent found a 33.8-minute charge (10 to 80%), against 42.6 minutes for Bayesian optimisation and 60 for the cell’s standard charge — staying inside the anode-plating and 45 °C limits. The results are simulated; no physical cell has been tested yet. Data-center hardware, robotics and power electronics are next, and the code is open source on GitHub.