About Liquid AI

Spun out of MIT CSAIL, we build general-purpose AI systems that run efficiently across deployment targets, from data center accelerators to on-device hardware, ensuring low latency, minimal memory usage, privacy, and reliability. We partner with enterprises across consumer electronics, automotive, life sciences, and financial services. We are scaling rapidly and need exceptional people to help us get there.

The Opportunity

At Liquid AI, we’re building novel model architectures that require GPU performance engineering beyond what standard frameworks can deliver.

You’ll design, implement, and optimize custom GPU kernels that make our models faster and more efficient across training, post-training, and inference. Working closely with researchers and engineers, you’ll identify performance bottlenecks, translate architectural innovations into efficient implementations, and deliver measurable improvements in production workloads.

This is a hands-on role for someone who enjoys understanding how GPUs actually execute code, from memory access patterns and warp scheduling to tensor core utilization. You’ll have the ownership to investigate complex performance problems, make informed engineering tradeoffs, and ship optimizations that directly impact how our models perform.

What We’re Looking For

We need someone who:

Reasons from first principles: You understand GPU architecture and can explain why a workload is compute-bound or memory-bound, how latency and throughput interact, and which optimizations are likely to matter.

Builds at the kernel level: You’re comfortable writing, debugging, and optimizing GPU kernels using CUDA or Triton. You understand the implementation details, not just the abstractions.

Optimizes systematically: You form hypotheses, use profiling and benchmarks to identify bottlenecks, evaluate tradeoffs, and validate improvements against real workloads.

Turns research into performance: You can translate novel model architectures and research ideas into efficient, correct, and maintainable implementations.

Owns problems end-to-end: You can take an ambiguous performance challenge from investigation through implementation, integration, benchmarking, and production deployment.

Communicates technical decisions clearly: You explain your reasoning, challenge assumptions constructively, and collaborate effectively with researchers and engineers.

The Work

Design, implement, and optimize high-performance GPU kernels for Liquid AI’s novel model architectures.

Profile training, post-training, and inference workloads to identify bottlenecks and improve latency, throughput, and hardware utilization.

Optimize memory access patterns, thread and block configurations, register usage, shared memory, and tensor core utilization.

Integrate custom kernels into PyTorch pipelines through custom operators, extensions, and appropriate dispatch mechanisms.

Evaluate numerical precision, correctness, and performance tradeoffs across different workloads and hardware configurations.

Build and maintain benchmarking infrastructure, correctness tests, and performance regression checks.

Collaborate with researchers to translate architectural innovations into production-ready GPU implementations.

Desired Experience

Must-have

Hands-on experience developing and optimizing custom GPU kernels using CUDA or Triton, beyond relying exclusively on high-level libraries.

Strong understanding of GPU architecture, including memory hierarchy, thread execution, warps, shared memory, register pressure, occupancy, and compute versus memory bandwidth limitations.

Demonstrated ability to reason about GPU performance, identify bottlenecks, and make informed optimization decisions.

Experience profiling, benchmarking, and debugging GPU workloads using tools such as Nsight Systems, Nsight Compute, or equivalent tooling.

Strong C++ programming skills and experience working with performance-sensitive code.

Nice-to-have

Experience with CUTLASS, tensor cores, or advanced matrix multiplication optimizations.