Senior AI Infrastructure Engineer - Model Training

Kodiak AI, Inc. (Nasdaq: KDK), founded in 2018, is a leader in Physical AI, developing driverless technology that powers machines that move. Its Kodiak Driver is a vehicle-agnostic autonomous driving system that combines AI-powered software with modular hardware. Today it operates in long-haul trucking, industrial trucking, and defense. In 2024, Kodiak launched commercial Driver-as-a-Service operations with no one in the cab, and plans to launch driverless long-haul service between Dallas and Houston by the end of 2026.

Kodiak has partnered with many of the biggest names in freight, including Bosch for scaled hardware production, AWS, and leading customers including Atlas Energy Solutions, IKEA, Bridgestone, Werner Enterprises, C.R. England, Roehl Transport, and General Dynamics Land Systems.

Kodiak's AI is only as good as the speed at which we can train it. Every improvement to our models – from GigaFusionNet to large-scale world models – depends on infrastructure that turns thousands of hours of multimodal driving data into training throughput. We are looking for engineers who make model training fast: streaming massive camera, LiDAR, and radar datasets without stalling a single GPU, sharding data and models efficiently across nodes, and extracting every FLOP from the latest hardware. If you measure your impact in tokens per second and GPU utilization, this role is for you.

In this role, you will:

• Design high-throughput data loading and streaming systems for multimodal sensor data (camera, LiDAR, radar), including dataset formats, sharding strategies, and prefetching pipelines that keep GPUs saturated

• Build and optimize distributed training infrastructure across multi-node GPU clusters, applying data, tensor, pipeline, and fully sharded (FSDP/ZeRO) parallelism to models that don't fit on a single device

• Maximize utilization of modern accelerators such as NVIDIA B200s through mixed-precision training (BF16/FP8), fused kernels, memory optimization, and communication/computation overlap

• Profile end-to-end training pipelines to find and eliminate bottlenecks across storage, network, CPU preprocessing, and GPU compute

• Develop scalable dataset construction pipelines that convert petabytes of raw driving logs into training-ready, streamable formats

• Partner with ML teams to scale new architectures from prototype to full-cluster training runs efficiently and reliably

What you’ll bring:

• BS, MS, or PhD in Computer Science or a related field, and at least 2-3 years of industry experience in ML systems or infrastructure

• Hands-on experience with distributed training frameworks and techniques (PyTorch DDP/FSDP, DeepSpeed, Megatron, NCCL) and a strong grasp of parallelism trade-offs