ML Systems Engineer — Inference Acceleration

Arago’s mission is to re-engineer the foundations of computing from first principles.

The explosive growth of AI is pushing the industry to rethink how processors are built. Arago is meeting that challenge with a proprietary technology that fuses optical and CMOS technologies to deliver an order-of-magnitude increase in performance.

Arago is the fastest, and currently the only, company to have built such a processor. It's backed by leading deep-tech investors and some of the most respected figures in semiconductors and computing, including the CEO of Arm, the founder of macOS who worked directly with Steve Jobs at Apple, an Nvidia Fellow, the Head of Optics at Google, and many other industry leaders.

Our work is guided by three clear values: do great things, move with high velocity, and operate as one unit. We work in a demanding environment where constant learning, ownership, and execution are expected, and where exceptional people have the opportunity to do their life’s work.

Optimize the execution and serving of modern AI models on Arago's custom accelerator. Work across kernels, model execution, multi-device distribution, runtime, and inference serving, while helping shape the software stack around the capabilities of Arago's hardware.

• →

  Analyze modern AI workloads and identify kernel-, runtime-, memory-, and system-level bottlenecks on Arago's accelerator.

• →

  Develop and optimize custom kernels, fused operators, and execution strategies to maximize device utilization.

• →

  Design efficient mappings of models and operators across multiple Arago devices, including communication and synchronization strategies.

• →

  Develop inference-serving techniques such as continuous batching, paged KV caches, prefix/context caching, chunked prefill, and prefill/decode interleaving or disaggregation.

• →

  Build profiling, benchmarking, and performance-analysis infrastructure spanning kernels, full models, and serving workloads.

• →

  Work closely with Arago's hardware, compiler, and runtime teams to co-design software abstractions and influence future hardware features based on real model workloads.

• Strong experience in high-performance ML inference, GPU/accelerator programming, or ML systems engineering.

• Deep understanding of computer architecture, accelerator/GPU execution models, memory hierarchies, parallelism, and performance bottlenecks.

• Experience developing and optimizing custom kernels using CUDA, Triton, ROCm/HIP, or equivalent low-level programming environments.

• Experience with operator fusion, tiling, scheduling, data movement optimization, graph execution, and profiling of compute- and memory-bound workloads.

• Strong understanding of distributed model execution, including tensor, pipeline, sequence, and/or expert parallelism and communication/computation overlap.

• Hands-on experience with modern inference-serving systems such as vLLM, SGLang, TensorRT-LLM, or equivalent, including KV-cache management, continuous batching, paged attention, and prefill/decode scheduling.

• Strong C++ and Python skills, and comfort working on a custom accelerator stack where compiler, runtime, kernels, and abstractions are actively being developed. Exposure to or experience with MLIR and MLIR dialects is a strong plus.

• Language: English at a proficient level.