Agentic RISC-V Verification Platform
Python · SystemVerilog · cocotb · UVM · coverage-driven verification
The problem
When an automated agent reports that a design is verified, why should anyone believe it? The interesting failure in verification is not the crash — it is the green result that means nothing.
AVA — the pipeline
Give it RTL and it runs semantic analysis, generates a testbench, executes the design against a golden ISS, compares the two commit logs instruction by instruction, analyses coverage, and generates new tests aimed at what the campaign has not yet reached. Twelve agents, most of which run with no EDA toolchain at all.
VOE — the research track underneath
VOE asks the harder question. It is built on a small frozen kernel in which a claim cannot exist without a witness — a real artifact, produced by a real tool, that can be re-checked later. Every policy change is pre-registered before data is collected and recorded in an append-only ledger, including the ones that failed.
What it buys
Coverage-guided test generation and automated bug-discovery workflows that accelerate validation cycles — and, more importantly, a verification result you can actually trust.