$2.7M · Embedd + 2 more

What they're building

Embedd is a board support package generator. You point it at a chip's datasheet and it produces the device drivers, the devicetree, and the BSP that let real software talk to that silicon. Targets include Zephyr, bare-metal, Linux and AUTOSAR. The primitive underneath is the digital twin: Embedd models the chip itself from its documentation, then runs AI agents against that model to emit integration code and the tests and simulations to check it. The distinction matters. Most code-generation tools point a language model at documentation and hope. Embedd's own framing is deterministic generation from a chip spec, which is the only version that survives contact with an automotive safety review. A driver that is 95% right is a driver that bricks a board in the field. The wedge is the semiconductor vendors themselves, not the device makers. Embedd went live commercially in April 2026 and is already working with Microchip Technology, where it is generating Zephyr ecosystem support. That is the sharper business: a chip vendor whose parts lack drivers for the popular RTOS loses design wins, and manually staffing that backlog across a catalogue of thousands of parts is not economic. Embedd sells the vendor its own ecosystem coverage. The company claims production-ready semiconductor software up to six times faster than hand-writing it. Treat the multiple as a vendor number, but the underlying observation is not controversial to anyone who has read 3,000 pages of a reference manual to make one I2C sensor work.

Why this matters

Every robot, drone, EV and infusion pump is a software stack sitting on dozens of chips that share no common language. The founders did not find this problem in a market map. They ran a hardware company, hit Covid-era chip shortages, had to source whatever parts were available, and rewrote their firmware from scratch each time. Then Russia invaded Ukraine and disrupted them again. The pain is autobiographical. The timing argument is the physical-AI buildout. Model capability is no longer the bottleneck for robotics; the bottleneck is the unglamorous integration layer between silicon and the stack above it, and that layer is still hand-written by scarce embedded engineers. Seedcamp led a pre-seed into a company selling picks and shovels to the shovel makers, at the exact moment demand for embedded talent outstrips supply.