Inference Infrastructure Engineer Intern

About the role

At Rhoda AI, we’re building the next generation of generalist intelligent robots. We own the full robotics stack from high-performance hardware and robot systems to the infrastructure and state-of-the-art foundation world models that control our robots. Our robots are designed to be generalists capable of operating in complex, real-world environments and handling long-tail edge cases, made possible by our cutting edge research and end-to-end system design. We've raised over $450M and are investing aggressively in model research, infrastructure, hardware development, and manufacturing scale-up to make generalist robotics a reality.

The Role

Our foundation world models have to run fast enough to control robots in the real world, and efficiently enough to support large-scale research. As an Inference Infrastructure Engineer Intern, you'll work with our infra and research engineers to make model inference faster and more efficient across the stack, from data pipelines down to GPU kernels. You'll own a scoped, high-impact project end to end, ship it into our production or research workflow, and present your results to the team.

What You'll Do

Profile and benchmark inference for large video and world models to find bottlenecks in latency, throughput, and memory.

Apply and evaluate inference optimization techniques like quantization and low-precision inference, step distillation and sampling acceleration for diffusion-style models, attention and caching optimizations across frames, model compilation, and graph-level optimizations.

Make data movement and processing more efficient: video decoding and preprocessing, input pipelines, host–device transfer, caching, and storage/IO paths that feed inference and training.

Write and tune GPU kernels (CUDA, Triton, or similar) for performance-critical operators, and integrate them into our inference stack.

Build tooling and benchmarks that let the team track performance regressions and improvements reliably.

Work with researchers to turn promising ideas into production-quality infrastructure.

What We're Looking For

You don't need all of these. Strong depth in any one of the areas below is enough to apply:

ML inference optimization: hands-on experience making deep learning models run faster or cheaper, e.g. quantization (INT8/FP8), pruning/distillation, operator fusion, mixed precision, or compilers and runtimes such as torch.compile, TensorRT, ONNX Runtime, or TVM.

Data efficiency optimization: experience speeding up data loading and preprocessing pipelines (especially video or other high-bandwidth sensor data), cutting IO and memory overhead, or improving data/compute utilization in ML systems.

GPU kernel optimization: experience writing or tuning GPU kernels in CUDA, Triton, CUTLASS, or similar; familiarity with the GPU memory hierarchy, occupancy, and profiling tools (Nsight Systems/Compute, PyTorch profiler).

Basic Qualifications

Currently pursuing a BS, MS, or PhD in Computer Science, Electrical/Computer Engineering, or a related field.

Strong programming skills in Python and/or C++.

Solid understanding of deep learning fundamentals and at least one framework (PyTorch preferred).

Comfortable reading performance data and reasoning about trade-offs between latency, throughput, memory, and accuracy.

Nice to Have

Experience with video model inference is a huge plus, e.g. optimizing video generation, video diffusion, or world models, or real-time video understanding pipelines.

Research experience in systems or ML, e.g. publications or substantial research projects (MLSys, OSDI/SOSP, CVPR/ICCV, NeurIPS/ICML/ICLR, or similar).

Experience with real-time or latency-critical inference, e.g. on robots or edge/embedded accelerators.

Experience with multi-GPU inference (tensor/sequence/pipeline parallelism).

Contributions to open-source ML systems projects.

What You'll Get

A real project with measurable impact on robots operating in the real world.

Mentorship from experienced infra and research engineers.

Access to modern GPU compute for experiments.