Staff AI Application Engineer - Heterogeneous Computing
Job description
Staff AI Application Engineer in the Model Enablement team (Application Software), building the model enablement platform and deciding how each model runs across the Metis or Europa AIPU, host CPU and host accelerators. The role combines architecture and hands-on work across the model frontend, graph compiler, runtime and application pipeline.
At a glance
- Work mode
- Remote · UK, Belgium, Italy
What you'll do
- Build the model enablement platform and agent harness: reusable tools to partition, compile, run and profile; environments with access to real hardware; closed-loop runs that apply a change, measure it against the reference and repeat; and verification (differential testing, regression detection) integrated into CI.
- Own model ingestion: how a PyTorch or ONNX model is captured, analysed and split across the AIPU's matrix and RISC-V vector units, the host CPU and the host's accelerators (GPU, NPU, video decoder), with placement tags that users can read and override; build this with the graph compiler team so it works for every model.
- Extend the split to large models, cascades and multi-model pipelines, for example with expert offloading and caching across device and host memory; agree the plan with the graph compiler team.
- With the application team (pipeline builder) and the runtime team, bring each model to its best end-to-end performance: overlap of host and device work, zero-copy buffers, batching, multi-core scheduling, and models with several inputs.
- Profile device time, host time and data movement, and confirm each speedup in the full pipeline.
- Study the models that matter most to customers, on Metis today and on Europa, and turn each lesson into a platform improvement.
- Use AI coding agents in daily work to build, test and verify your own work.
What we're looking for
- 8+ years of engineering experience in systems, compilers, runtimes or performance, including system-level design ownership.
- Hands-on experience with heterogeneous computing across CPU, GPU, DSP, NPU or FPGA, including reasoning about compute, memory bandwidth, latency and data movement.
- Solid fundamentals in computer architecture, operating systems and parallel programming.
- Familiarity with ML graph representations and compiler concepts: graph capture (torch.export, FX, ONNX), lowering, operator fusion, partitioning, memory planning.
- Strong C++ and Python.
- Daily use of AI coding agents on real engineering work, with the judgment to review and test their output.
Nice to have
- Hands-on experience with MLIR or another compiler framework (TVM, XLA, TorchInductor, IREE).
- Experience with an inference runtime or execution engine (ONNX Runtime execution providers, a custom NPU or DSP runtime).
- Experience with GPU compute (CUDA, OpenCL, Vulkan) or SIMD optimization on Arm or x86.
- Experience with large-model execution techniques such as offloading, expert caching, speculative decoding, pipelining across devices.
- Experience building agent environments, evaluation harnesses or verification frameworks (differential testing, regression detection), or MCP-based tooling.
Your team
Model Enablement team (Application Software)
The team builds the model enablement platform; during model bring-up, the role covers the frontend and graph side while a device-side engineer covers the backend.
Apply
Direct apply at Axelera AI
Opens the job posting at Axelera AI
About Axelera AI
Creating the next-generation AI platform
Axelera AI is a deep-tech company creating a next-generation AI platform. It has launched its Metis™ AI Platform, which achieves a 3-5x increase in efficiency and performance, and has visibility into a strong business pipeline exceeding $100 million.
Mission
To support anyone who wants to help advancing humanity and improve the world around us.
Key facts
- AI hardware and software
- Launched the Metis™ AI Platform, which achieves a 3-5x increase in efficiency and performance
- Visibility into a strong business pipeline exceeding $100 million
- Global industry pioneer