About the role
As a Research Engineer on the Frontier Foundations team, you will partner with Research Scientists to make large-scale model training and inference as fast and efficient as the hardware allows. You will work at the lowest level of our stack, writing and tuning GPU kernels and removing the memory and communication bottlenecks that limit large Mixture-of-Experts models on the latest GPUs.
Building on our track record of AI-powered solutions (e.g., Bits AI, Bits Evolve, and our time series foundation model Toto), Datadog AI Research (DAIR) tackles high-risk, high-reward problems grounded in real-world challenges in cloud observability and security.
Frontier Foundations builds and maintains the high-performance infrastructure that lets us train and serve large language models at scale. It supports both of our research areas:
Trained Agents for Observability - Post-training models to operate autonomously across Datadog's domain. SRE incident response is our first target, with a clear path to code repair, security response, and infrastructure optimization. We build the simulation environments, RL training loops, and evaluation infrastructure needed to train agents that match or surpass frontier models at a fraction of the cost.
What You'll Do
- Write, tune, and debug GPU kernels (CUDA, Triton, or similar) for the operations that dominate LLM training and inference.
- Optimize for the latest NVIDIA GPU generations, including memory hierarchy, tensor cores, and low-precision formats such as FP8 and FP4
- Reduce communication cost in distributed workloads (all-to-all, all-reduce) and overlap communication with compute across NVLink and across nodes
- Profile end-to-end training and rollout workloads to identify whether they are compute-, memory-, or communication-bound, and quantify the headroom
- Build benchmarks, profiling tools, and regression tests that guide optimization decisions
- Collaborate with Research Scientists to turn new model architectures into efficient implementations
- Explore emerging techniques such as low-precision training and serving, speculative decoding, and fused or persistent kernels
Who You Are
- You have depth in at least one of: GPU kernel development and optimization (CUDA or similar), low-latency communication and collective operations (e.g., RDMA, NCCL), or large-scale training infrastructure
- You can reason about hardware constraints (memory, bandwidth, interconnect topology) and explain where a bottleneck is and how you would confirm it