Research Engineer Intern - AI Systems

Yotta Labs

6d•Remote

About The Position

Yotta Labs is building the next generation multi-silicon AI cloud and runtime platform to power the world’s most demanding AI workloads. We enable training and inference across NVIDIA GPUs, AMD GPUs, and AWS Trainium, helping AI companies achieve the best performance and economics across heterogeneous hardware. Our mission is to provide high-performance AI computing and Model API services, enabling AI companies, research labs, and enterprises to train, deploy and integrate cutting-edge models at scale.

Requirements

Currently pursuing a BS, MS, or PhD in Computer Science, Computer Engineering, or a related field.
Solid programming skills in Python and familiarity with C++.
Understanding of GPU/accelerator architecture fundamentals (memory hierarchy, parallelism, occupancy) from coursework, research, or projects.
Experience writing CUDA, Triton, ROCm/HIP, or Neuron kernels — class projects and personal projects count.
Strong understanding of AI frameworks (e.g., PyTorch, Dynamo, LMCache), model architectures and profiling tools (e.g. Nsight, ROCm Profiler, or Neuron Profiler).
Strong problem-solving skills and the ability to work independently in a collaborative, remote environment.

Nice To Haves

Contributions to open-source AI infra projects like vLLM, SGLang, PyTorch, or Triton.
Familiarity with LLM inference internals — FlashAttention, PagedAttention, continuous batching, speculative decoding, MoE, or quantization.
Experience with profiling tools (e.g. Nsight, ROCm Profiler, Neuron Profiler, or PyTorch Profiler) and performance debugging on real workloads.
Publications in top-tier conferences like MLSys, OSDI, SOSP, NSDI, SC, HPCA, or ISCA

Responsibilities

Implement and optimize compute kernels for Attention, GEMM, MoE, and quantization on NVIDIA, AMD, or AWS Trainium.
Build custom operators using CUDA, Triton, ROCm/HIP, or the Neuron SDK with PyTorch/XLA.
Profile and improve inference performance in vLLM, SGLang, and our custom runtimes — kernel fusion, scheduling, KV-cache and memory optimizations.
Build benchmarks, chase down performance regressions, and turn profiler traces into concrete speedups.
Ship code upstream to open-source AI infrastructure projects, with tests and documentation.