Senior / Staff Research Engineer, Generative AI Deployment

• Be part of a multidisciplinary team of Research Scientists and Engineers building a best-in-class multi-sensor simulation stack.

• Take cutting-edge models from research to production. Turn research code and checkpoints into versioned, reproducible, exportable production models. These models run in closed-loop simulation next to GPU renderers and the autonomy stack, and in the pipelines that reconstruct 4D simulation scenes.

• Make generative models leaner and faster to run. Profile end to end, from the model graph to GPU kernels and memory movement. Apply techniques like model compilation (e.g., TensorRT), mixed precision and quantization, batching, CUDA graphs, and custom kernels to achieve latency and memory budgets. For productionizing diffusion and flow-matching models, this can also include domain-specific techniques like faster sampling techniques and distillation.

• Own correctness as part of performance. Build benchmarks and numerical-parity tests focused on simulation realism and downstream autonomy metrics. These checks make sure that each speedup keeps simulation realism intact, and they catch regressions in CI.

• Build shared deployment tooling. Design the export, versioning, model registry, and build tooling that lets the sensor simulation team ship new models to simulation with confidence.

• Shape models for deployment. Work with research scientists early, so that new architectures are exportable, efficient, and hardware-friendly from the get-go.

Qualifications:

• Shipping Production Software. You have experience reading and developing production-quality software at scale, not only prototypes. You have shipped ML systems that other engineers and researchers depend on every day.

• Model Optimization and Deployment. You have taken models from a research codebase to an optimized production artifact, and measured both the speedup and the accuracy change. You have hands-on experience with several of the following: TensorRT, torch.compile, torch.export, ONNX, post-training quantization, quantization-aware training, and pruning. You have handled models with custom CUDA operators or other operations that don't export cleanly.

• GPU Performance. You profile with tools such as Nsight Systems, Nsight Compute, and the PyTorch profiler. You reason about memory bandwidth, kernel launch overhead, and host-device synchronization. You can write or modify CUDA or Triton kernels.

• Modern ML Architectures. You can read research code for modern architectures, such as vision transformers, diffusion and flow-matching models, and vision foundation models, and see where the compute and memory go. You know which accuracy trade-offs matter when you lower precision, compile a model, or cut its sampling steps.

• Cloud Infrastructure. You have experience launching and debugging GPU jobs in the cloud (AWS, GCP, or equivalent). This includes job orchestration, containerized environments, and keeping CUDA, PyTorch, and TensorRT versions compatible across a fleet.