Key takeaways ▼
Sirish Munipalli is Co-Founder and CTO of Moores Lab AI, where he leads the development of agentic AI for chip design verification, including SpecAgent(tm), the company’s specification and modeling platform. He has spent about 15 years designing and verifying complex silicon: FPGA prototyping at Intel, post-silicon validation at Qualcomm, and design verification and architecture modeling, including a vision processing block, for machine learning chips at Microsoft. A large part of that career has been in neuromorphic computing, where he worked on the architecture, design and verification of spiking neural network processors, built the software simulators and tooling around them, and converted deep learning models to run as spiking networks. He pairs that industry experience with a graduate research background in neuromorphic architectures and applied deep learning. He is a co-inventor on three pending patent applications covering AI-driven coverage closure and graph neural network acceleration.
What brought you to Moores Lab AI?
I have spent my career building and checking complex chips: FPGA prototyping at Intel, post-silicon validation at Qualcomm, and design verification and architecture modeling of a vision block on machine learning chips at Microsoft. In between, I spent years in neuromorphic computing, working on the architecture of spiking neural network processors from the sensor interface through to the software simulator. Designing an architecture that new, with no established playbook, taught me how much depends on getting intent across from architect to designer to verifier. At every stop the same thing was true. The design was rarely the bottleneck. Proving it correct was. I watched skilled engineers spend their days translating specifications into testbenches, chasing failures through logs, and closing the last few coverage holes by hand. When language models became capable of reasoning over specifications and code together, it was clear this work could be rebuilt, not just patched. Shelly Henry, Shashank Chaurasia and I founded Moores Lab AI in 2025 to do that.
What is the biggest challenge your customer is facing?
Verification has outgrown the teams doing it. Industry surveys have long put it at more than half of project time, and the 20 24 Wilson Research study found that only 14% of projects achieved first-silicon success. Our customers are building more complex chips on shorter schedules, and they cannot hire verification engineers fast enough to keep up.
Underneath that is a quieter problem: context loss. An architect’s intent lives in a specification, a designer’s assumptions live in RTL, and the verification engineer has to reconstruct both. Every handoff loses something, and what gets lost often shows up later as a bug that escaped.
Why is your approach to solving this problem unique?
We build agents that carry out the verification workflow, not assistants that hand an engineer a code snippet. VerifAgent(tm) starts from the specification and produces a working UVM environment. DebugAgent(tm) traces failures to their cause. CoverageAgent(tm) works on closing the gaps that remain. Each one runs against real simulation, because simulation is the source of truth. A model’s opinion about whether a design is correct is worth nothing until a simulator agrees.
The foundation is SpecAgent, which I am leading. It turns a team’s source material, such as datasheets, design notes and addenda, into a single living specification where every statement traces back to its source and gaps are flagged. That becomes a unified knowledge base that architects, designers, verification engineers and our other agents all work from, so a change in the spec reaches everyone instead of getting lost in a handoff. We are extending it to generate C and SystemC models from the same specification, which gives teams a reference model for verification and a software virtual platform for early system and software development.
Throughout, the engineer stays in control, with every step visible and reviewable, and we fit into the simulators and flows a team already uses.
What other approaches are being used to solve this problem?
There are three that we see most often. The first is adding people or outsourcing, which works until the schedule or the budget runs out. The second is general-purpose AI coding assistants. They are impressive on software, but they do not know a team’s specification, methodology or simulator, and they miss the intricacies that make hardware bugs expensive. The third is AI features added to existing EDA tools, which improve individual steps but leave the workflow between steps untouched.
Formal methods and portable stimulus remain valuable alongside all of these,
which is why we partner with Breker.
What challenges will your customers face in the near future?
The first is volume. More companies are designing their own silicon, especially for AI workloads, and multi-die and chiplet designs multiply the interfaces that have to be verified. The talent pool is not growing at that rate. The second is new. As AI starts generating RTL, the amount of design that needs checking will grow faster than ever, and verification becomes the control point for trust. Teams will need to show how something was verified, by which agent, against which requirement, with evidence they can audit. The third is data. Customers want the benefit of AI without their design IP leaving their walls, so deployment and security will decide which tools they can use at all.
How has AI changed your business strategy?
We were founded on AI, so the real question is how its pace shapes our decisions. Frontier models improve every few months. A company whose value is a model will be overtaken by the next release. So we invest in what does not get commoditized: the workflows, the verification methodology, the integration with simulation, and the feedback loops that tell an agent it was wrong. We stay model-agnostic and expect to swap what sits underneath. It has also changed how we deliver. Customers buying into a new way of working want outcomes, not licenses alone, so our engineers work alongside theirs and we offer verification as a service. And we use our own agents internally, which is the fastest way to find out where they fall short.
Also Read:
From Frontend to the Frontiers of Silicon with Moores Lab AI
Focusing on the AI Big Picture with Moores Lab AI at DAC 2026
MooresLabAI at DAC 2026: Why the Future of Semiconductor Engineering Is Agentic, Not Just Generative
Share this post via: