Kandou AI | When AI Meets Semiconductor Verification to Accelerate Innovation

The Birth of a Beehive

Kandou AI story began around 2010 in ALGO, an EPFL laboratory led by Dr. Amin Shokrollahi. The challenge was straightforward yet critical: processors were becoming exponentially faster, but moving data between chips had become a major bottleneck. The research team explored a novel approach, making multiple electrical signals work together rather than treating each wire independently. This breakthrough led to Chord™ signaling and Chordal Codes, technologies designed to deliver higher data throughput while reducing power consumption, physical connections, and footprint.

Kandou was co-founded in 2011 as an EPFL spin-off by Amin Shokrollahi. The name “Kandou” means “beehive” in Farsi, Amin’s native language. This choice reflects the company’s core philosophy: just as individual bees work together to create something far greater than any single bee could achieve, Kandou’s technology makes multiple signals work together as a system. The name captures the company’s philosophy: intelligent cooperation between individual elements can create a much more powerful system.

The Verification Bottleneck

Despite Kandou’s success in high-speed connectivity, the company faced a universal challenge in semiconductor development: functional verification which consumes more than half of a project’s engineering effort. Translating complex design specifications into SystemVerilog Assertions (SVAs) remains largely manual, requiring scarce specialist expertise and creating risks of errors, late bug discovery, and longer time-to-market.

Existing Electronic Design Automation (EDA) tools can verify properties once they are written but cannot automatically understand complex specifications and derive the required assertions. This gap motivated Kandou to seek solutions beyond conventional semiconductor verification, leading to a collaboration with IDIAP’s Neuro-symbolic AI Group, which brings leading expertise in large language models, symbolic reasoning, auto-formalization, and LLM–solver refinement.

The SPECTRA Project

The Innosuisse SPECTRA project combines Kandou’s semiconductor knowledge with IDIAP’s research to automatically transform engineering specifications—including natural language, tables, and diagrams—into formal verification properties such as SystemVerilog Assertions. The key innovation lies in a multi-agent, neuro-symbolic architecture rather than relying on a large language model alone. Specialized AI agents interpret specifications, identify ambiguities and boundary conditions, generate assertions, and interact with simulation and formal verification tools. Candidate assertions are automatically checked, with errors and counterexamples fed back to the AI for iterative correction.

This creates a closed loop between generative AI and rigorous formal verification. The objective is not simply to generate plausible code, but to produce assertions that are formally checked, traceable to the original specification, and understandable by engineers—reducing manual effort while improving verification coverage, consistency, and product quality.

Alliance Mentoring, a Strategic Partnership

Kandou learned about Alliance through IDIAP. When preparing their Innosuisse proposal for the SME call, the team contacted Robert van Kommer, Alliance’s ICT and AI innovation mentor, relatively close to the submission deadline. Robert van Kommer responded quickly and agreed to support the team. He first reviewed key draft sections of the proposal—Solution, Value Creation, and Risks—providing valuable feedback that the team incorporated before uploading the updated proposal through the Innolink platform. He then conducted a second review of the full pre-final proposal on Innolink, identifying additional areas for improvement and potential issues, and offering practical suggestions to address them before submission.