Inference Optimization Engineer

Hedra•San Francisco, CA

20h•Onsite

About The Position

Hedra is the platform, models, and infrastructure for visual intelligence. We build models and systems that push the frontier of visual intelligence, along with the infrastructure required to make those models fast, efficient, reliable, and accessible at scale. We’re a team in San Francisco, backed by a16z and other leading investors. Researchers and engineers at Hedra work closely across boundaries, own problems end to end, and have significant influence over both what we build and how we build it. We’re looking for an Inference Optimization Engineer to work alongside our research team on making state-of-the-art visual models fast and efficient at inference time. You’ll work at the boundary between research and systems, taking new model architectures and figuring out how to run them efficiently on modern hardware. That means understanding where time and memory are being spent, identifying opportunities for algorithmic and systems-level improvements, and implementing optimizations across model architecture, inference algorithms, runtimes, kernels, and distributed execution. The problems rarely live neatly within one layer of the stack. Depending on what you find, you might modify how a model executes, develop a new inference technique, write a custom GPU kernel, rethink memory movement, or change how work is distributed across accelerators. We care more about technical depth, curiosity, and demonstrated ability than years of experience. We’re open to experienced ML systems engineers as well as exceptional early-career engineers or researchers who have already gone unusually deep on model performance, GPU systems, or efficient inference.

Requirements

Deep technical ability in efficient ML inference, ML systems, GPU computing, or adjacent research, demonstrated through research, production engineering, open-source contributions, or unusually ambitious independent work.
Strong understanding of how modern deep learning models execute on hardware, including the relationship between compute, memory, communication, and performance.
Experience profiling ML workloads, identifying bottlenecks, forming hypotheses, and driving measurable performance improvements.
Strong programming fundamentals in Python, C++, or another systems-oriented language.
Experience with some combination of PyTorch, CUDA, Triton, TensorRT, vLLM, SGLang, or comparable technologies.
Ability to reason across abstraction layers rather than treating model architecture, framework, runtime, kernel, and hardware boundaries as fixed.
Strong intuition for performance tradeoffs across latency, throughput, memory, numerical precision, model quality, and complexity.
Curiosity about how models work internally and a willingness to modify or rethink existing approaches when the performance problem calls for it.
Comfort working on ambiguous problems where the bottleneck, and sometimes even the right question, is not known in advance.
Ability to communicate technical ideas clearly and collaborate closely with research scientists and engineers.
Exceptional depth in one or more relevant areas, combined with the ability and curiosity to reason across the others, matters more to us than checking every box.

Nice To Haves

Experience optimizing large generative, multimodal, vision, or video models.
CUDA, Triton, CUTLASS, or other GPU kernel development.
Deep knowledge of GPU architecture, memory hierarchy, and hardware-aware optimization.
Experience with attention optimization, kernel fusion, memory-efficient execution, or custom operators.
Model compilation or graph optimization experience.
Quantization, sparsity, caching, speculative execution, or other efficient inference techniques.
Experience optimizing diffusion, autoregressive, transformer, or other large generative architectures.
Multi-GPU or multi-node model execution, including tensor, pipeline, sequence, or other forms of parallelism.
Experience optimizing communication or data movement between accelerators.
Experience with profiling tools such as Nsight Systems or Nsight Compute.
Contributions to ML systems, inference runtimes, compilers, GPU libraries, or performance-focused open-source projects.
Research or publications in efficient ML, ML systems, GPU computing, compilers, or related areas.

Responsibilities

Work directly with research scientists and engineers to make new visual models fast and efficient at inference time.
Profile model architectures and workloads to understand bottlenecks across compute, memory, communication, and model execution.
Develop and implement new approaches to improving inference latency, throughput, memory efficiency, and GPU utilization.
Explore algorithmic optimizations including quantization, sparsity, caching, compilation, attention optimizations, and alternative execution strategies.
Build or optimize GPU kernels using CUDA, Triton, or similar technologies when existing implementations leave performance on the table.
Optimize model execution across single-GPU, multi-GPU, and multi-node environments.
Reason about the interaction between model architecture and hardware, and work with researchers when architectural changes can unlock meaningful performance improvements.
Investigate communication, memory movement, parallelism, and distributed execution strategies for large visual models.
Build rigorous benchmarking, profiling, and performance-regression infrastructure to understand performance and evaluate new optimization ideas.
Evaluate new inference runtimes, compilers, frameworks, optimization techniques, and accelerator hardware.
Stay close to advances in efficient inference, GPU programming, model architectures, compilers, and ML systems research, and rapidly test promising ideas.
Help turn research breakthroughs into models that can be deployed and served efficiently at scale.