Lemurian Labs is building a hardware-agnostic, software-first AI infrastructure stack designed to allow AI workloads to run efficiently across CPUs, GPUs, TPUs, and accelerators from different vendors without hardware-specific rewrites. The company’s technology includes its Tachyon compiler and runtime, along with a proprietary logarithmic data type aimed at improving computational efficiency and memory usage. Pulse 2.0 interviewed Lemurian Labs Founder and CEO Jay Dawani to learn more.
Jay Dawani’s Background
When asked about his background and the experiences that eventually led him to Lemurian Labs, Dawani shared:
I grew up in Karachi, Pakistan, in a multi-generational household: lots of movement, lots of conversation, and dial-up internet until I was about 14.
School and I never really clicked; the system was built around memorization, and I was more interested in exploring ideas on my own terms. That pushed me to spend a lot of time by myself, which turned out to be a gift, allowing me to explore without limits, immersing myself in books, comics, and strategy games.
I swam competitively from age seven, and a second-place finish around age 13 taught me something foundational: I didn’t just enjoy winning; I really disliked losing. That shaped my discipline.
My first startup, in 2018, was an AI-powered e-commerce platform for immersive clothing shopping. It was a good idea at the wrong moment, since mobile hardware couldn’t support what we needed.
When I wound it down, a conversation with an investor led to me becoming interim CTO of a publicly traded blockchain gaming company at 24, after the board removed the founders. Those six months of restructuring and public-market scrutiny taught me more than most years before or since. That’s also when my hair started going grey.
Before starting Lemurian, I advised NASA’s Mars Rover program at Geometric Energy on vision-based navigation and planetary mapping. I co-founded Lemurian Labs in 2022 with Dr. Vassil Dimitrov.
How Lemurian Labs Started
When asked how the idea for Lemurian Labs came together, Dawani explained:
Back in 2019, I was leading research into collaborative robotics, warehouse robots, construction robots, drone swarms, trying to get multiple autonomous agents to coordinate in real time.
We built a hierarchical reinforcement learning framework that worked surprisingly well in simulation. Then we tried it on real robots and everything broke.
That sim-to-real gap was brutal, and even after finding a workaround, we only hit about 40-50% reliability.
That project convinced me the real bottleneck wasn’t the learning algorithms. It was the infrastructure underneath: world modeling, compute architecture, programming models. That’s the insight that led directly to founding Lemurian.
As CEO, I lead our mission to rebuild AI infrastructure. Originally we were thinking chip architecture, and we’ve since moved into compilers and runtime software, for better efficiency, accessibility, and performance across all kinds of hardware.
Navigating Strategic Pivots
When asked about his favorite memory from working at Lemurian Labs, Dawani reflected:
Honestly, this is one I haven’t answered directly in public yet, so I don’t want to put words in my own mouth here.
What I can say is that pushing through our two major strategic pivots, and proving people wrong who told us a startup couldn’t tackle this problem, is something I look back on with real pride.
If you want my actual favorite memory, you’d need to ask me that one directly.
Tachyon And The AI Infrastructure Stack
When asked about Lemurian Labs’ core products and features, Dawani detailed:
We’re building a hardware-agnostic, software-first AI infrastructure stack: a compiler and runtime that lets AI workloads run efficiently across CPUs, GPUs, TPUs, and accelerators from any vendor, whether that’s on-prem, cloud, or edge, without hardware-specific rewrites.
Our flagship compiler is called Tachyon, which we’re planning to launch through a beta testing program.
The stack we’ve built from the ground up includes a new compiler and runtime that turns clean, composable code into highly optimized, scalable programs that run seamlessly across CPUs, GPUs, and accelerators, eliminating the need for boilerplate, kernel code, or hardware-specific rewrites.
It ingests PyTorch models and runs them efficiently across various hardware platforms, enabling seamless model training and inference.
Additionally, we’ve developed a novel logarithmic data type that improves computational efficiency and memory usage by providing a smaller, more precise representation than traditional floating-point formats.
I think of the whole platform as giving researchers a unified programming and execution environment. Just as Stripe abstracted away payments complexity, we’re abstracting away heterogeneous AI compute complexity.
The Sim-To-Real Challenge
When asked about some of the technical and strategic challenges the company has faced, Dawani noted:
The sim-to-real gap I mentioned was probably the sharpest technical challenge.
The more detailed and accurate we made our simulations, the worse our policies performed in reality, because the agents were learning to exploit tiny quirks of the simulator.
The counterintuitive fix was using lower-fidelity simulation with more randomness, which forced the agents to learn strategies that actually generalized.
At the company level, we’ve had to navigate two major strategic pivots: from robotics, to data-center chip design, to where we are now, solving the software layer for heterogeneous systems.
Each pivot meant letting go of work we were deeply invested in and starting over, often while people told us it couldn’t be done.
Technology Evolution
When discussing how Lemurian Labs’ technology has evolved since launching, Dawani explained:
We started with a bold, hardware-first thesis: build the entire foundational stack for autonomous, software-first machines, from chip architecture up through the compiler and programming model.
Most people told us that was impossible or impractical.
We pivoted from robotics into designing chips for the data center, and eventually realized the deeper problem wasn’t the hardware at all. It was how developers will program heterogeneous systems at scale.
That’s what pushed us toward the software-centric, compiler-and-runtime approach we have today, rather than custom silicon.
Major Company Milestones
When asked about some of Lemurian Labs’ most significant milestones, Dawani highlighted:
Our major milestones include founding Lemurian in 2022, working through two strategic pivots, and then closing a $28 million oversubscribed Series A round in December 2025, co-led by Pebblebed Ventures and Hexagon.
Oval Park Capital, which had led our seed round three years earlier, joined alongside more than a half dozen other backers.
More recently, in March 2026, we strengthened our leadership team: Java pioneer Kim Polese joined our board, and MIT compiler leader Saman Amarasinghe came on as a Technical Advisor.
Supply Chain Planning
When asked about a specific customer success story, Dawani shared:
One example I can point to: we adapted some of our multi-agent AI research to help a company improve its supply chain planning.
They had multiple facilities with different capacities and lead times, and they’d been making planning decisions in spreadsheets, which made trade-offs hard to see and slow to respond to.
We modeled each facility as its own agent and trained the system to coordinate decisions across the network.
Instead of one “optimal” plan, they could explore scenarios, prioritizing speed over cost in one case, resilience over efficiency in another.
It shifted them from reacting to problems to proactively designing around them.
I haven’t named that customer publicly, so if you need specifics, that’s something to follow up on directly with us.
$28 Million Series A
When asked about Lemurian Labs’ funding, Dawani said:
We closed an oversubscribed $28 million Series A in December 2025, co-led by Pebblebed Ventures and Hexagon.
Oval Park Capital, which had led our seed round three years earlier, joined alongside more than a half dozen other backers.
Competitive Differentiation
When asked what differentiates Lemurian Labs from its competition, Dawani emphasized:
My philosophy is to avoid competing head-on with the foundation model companies.
AI has natural monopoly dynamics, and if your business breaks the moment an API changes, you don’t really have a business.
So we deliberately chose a problem where size is a moat rather than a liability: rethinking the infrastructure underneath autonomous agents requires long time horizons and deep technical insight, which most companies won’t commit to.
We’re not tied to any single chip vendor, and I think our combination of a custom compiler, runtime, and proprietary logarithmic number format sets us apart from others working in adjacent spaces.
Future Goals
When discussing Lemurian Labs’ future goals, Dawani emphasized:
I want us to build the infrastructure for truly autonomous agents: systems that understand context, pursue goals over time, and learn from experience in the real world, rather than the train-once, deploy-frozen model most AI runs on today.
I want our platform to make compute fungible, so researchers can iterate on new agent architectures without rewriting kernels, and companies can deploy at scale without locking into one vendor’s hardware roadmap.
Near-term, we’re using our Series A funding to expand our engineering team, accelerate product development, and deepen partnerships focused on sustainable compute, with Tachyon heading toward a beta launch.
Building For The Future Of AI
When invited to discuss other ideas shaping Lemurian Labs’ approach to AI infrastructure, Dawani concluded:
I’d point to the five things I think anyone shaping AI’s future needs to understand: scale alone isn’t enough, learning from interaction is the right direction even though today’s methods are inefficient, today’s “agents” aren’t really agents yet, we need world models that let systems mentally simulate reality, and our current infrastructure wasn’t built for continuously learning systems.
That last one is really the thesis behind everything we’re building at Lemurian.
My advice to other AI founders is simple: don’t build something that breaks the moment an API changes.
Build on proprietary data, deep domain expertise, ownership of a workflow, or distribution nobody else has, and be able to answer honestly what you can do that OpenAI, Google, or Anthropic can’t or won’t do.
I keep coming back to a Marcus Aurelius line, “The impediment to action advances action”, because in my experience, progress doesn’t happen after you clear an obstacle; it happens through it.
