Senior AI Application Engineer - LLM & VLM Enablement

Axelera AI; United States; À distance; USD 150 000 - 230 000; Plein temps

Avantages offerts par ce poste

Pension plan

Employee insurances

Company shares

Résumé du poste

Axelera AI is seeking a Senior AI Application Engineer to own LLM/VLM support on Metis and Europa AIPUs, from studying new models to shipping them in the Voyager SDK.

The role blends ML inference with software engineering, using AI coding agents to build, test and verify work, and shaping GenAI capabilities for edge devices within a dynamic, international team.

Qualifications

• 5+ years of software engineering experience with ML inference, ML systems or ML tooling.
• Solid understanding of transformer inference: attention variants, KV caching, prefill vs decode, sampling, tokenization, and how a VLM joins a vision encoder to a language model.
• Working knowledge of LLM quantization (weight-only and weight-activation schemes, calibration, outlier handling).
• Strong Python and PyTorch, and experience shipping production-quality software.
• Daily use of AI coding agents on real codebases, with good judgment about when to verify or reject their output.
• Ability to turn research code and hardware constraints into a practical design.

Responsabilités

• Study new LLM and VLM families and architectures, prototype and profile them on Metis and Europa, and decide how they should run on our hardware.
• Turn results into Voyager SDK features with stable APIs, tests, documentation and release gates.
• Own the PyTorch reference with quantization baked in for each LLM/VLM block, weight mapping, and a contract for the compiler team.
• Bring LLM and VLM serving into the Voyager SDK pipeline builder: prefill and decode, KV-cache management, streaming text output and multimodal inputs.
• Work with regression and quantization owners to measure accuracy and explain gaps.
• Extend the model enablement platform for bring-up: environments with access to compiler, runtime and hardware; closed-loop runs to measure accuracy and speed.
• Recommend what the Voyager SDK needs for GenAI, agents and speech on edge devices.

Connaissances

Software engineering

Transformer inference

Python

PyTorch

AI coding agents

System design

Outils

llama.cpp

vLLM

ONNX Runtime GenAI

ExecuTorch

Description du poste

About Us

Axelera AI is not your regular deep-tech company. We are creating the next-generation AI platform to support anyone who wants to help advancing humanity and improve the world around us.

In just five years, we have raised a total of $450 million and have built a world-class team of 250+ employees (including 60+ PhDs with more than 40,000 citations), both remotely from 21 different countries and with offices in Belgium, France, Switzerland, Italy, the UK, headquartered at the High Tech Campus in Eindhoven, Netherlands.

We have also launched our Metis™ AI Platform, which achieves a 3-5x increase in efficiency and performance, and have visibility into a strong business pipeline exceeding $100 million.

Our unwavering commitment to innovation has firmly established us as a global industry pioneer.

Are you up for the challenge?

Position Overview

As a Senior AI Application Engineer in the Model Enablement team (Application Software), you will own LLM and VLM support on Metis and Europa AIPUs, from studying new models to shipping them in the Voyager SDK.

The role combines LLM/VLM knowledge with solid software engineering. Our team uses AI coding agents in daily work, and you will use them to build, test and verify your own work.

We also welcome experienced field application engineers, technical leads and engineering managers who want to return to hands‑on engineering. Experience in planning, delegating and reviewing work transfers well to working with AI agents.

Key responsibilities

• LLM/VLM study: Study new LLM and VLM families and architectures (for example mixture-of-experts, small VLMs, new attention and decoding schemes). Prototype and profile them on Metis and Europa, and decide how they should run on our hardware.

• Productization: Turn your results and our ML research group's results into supported Voyager SDK features, with stable APIs, tests, documentation and release gates.

• Model-side contracts: For each LLM/VLM building block, own the PyTorch reference with quantization baked in, the weight mapping, and one contract that lists every input, output, weight and persistent state (such as a paged KV cache), and hand that contract to the compiler team. The runtime and tests use the same definition.

• Pipeline integration: Bring LLM and VLM serving into the Voyager SDK pipeline builder: prefill and decode, KV-cache management, streaming text output and multimodal inputs.

• Quality: With the regression and quantization owners, measure accuracy against the floating-point reference, and close or explain each gap.

• LLM/VLM bring-up platform: Extend the model enablement platform for LLM and VLM bring-up: environments with access to the compiler, runtime and real hardware (for example through MCP tools), closed-loop runs that apply a change, measure accuracy and speed against the reference and repeat, and evaluations that score each run.

• GenAI direction: Recommend what the Voyager SDK needs for GenAI, agents and speech on edge devices, and which model zoo demos show it best.

Qualifications

Essential

• 5+ years of software engineering experience, with at least 2 years on ML inference, ML systems or ML tooling.

• Solid understanding of transformer inference: attention variants, KV caching, prefill versus decode, sampling, tokenization, and how a VLM joins a vision encoder to a language model.

• Working knowledge of LLM quantization (weight-only and weight-activation schemes, calibration, outlier handling), and the ability to tell a quantization loss from an implementation bug.

• Strong Python and PyTorch, and experience shipping production-quality software.

• Daily use of AI coding agents on real codebases, with good judgment about when to verify or reject their output.

• Ability to turn research code and hardware constraints into a practical design.

Preferred

• Experience with inference engines or serving stacks (for example llama.cpp, vLLM, ONNX Runtime GenAI, ExecuTorch).

• Experience with kernel development or a tensor compiler (for example Triton, TVM or MLIR), or with profiling models on accelerator hardware.

• Experience with speech (ASR, TTS) or agentic applications on embedded devices.

• Experience building evaluation or testing harnesses for ML systems or AI agents, or tool-calling agents (for example MCP servers).

Location

We offer a flexible working arrangement, with options to:

• Work from one of our Axelera AI offices (Leuven in Belgium, Amsterdam and Eindhoven in the Netherlands, Zurich in Switzerland, Florence and Milan in Italy or Bristol in the United Kingdom) if you're already based in the vicinity.

• Work fully remotely from any European country (incl. the UK) you are already in.

• Relocate with us and work from Italy (Florence or Milan) or the Netherlands (Amsterdam or Eindhoven).

Kindly note that priority will be given to candidates who are [interested in being] based in Belgium or Italy.

What weoffer

This is your chance to shape and be part of a dynamic, fast-growing, international organization. We offer an attractive compensation package, including a pension plan, extensive employee insurances and the option to get company shares.

An open culture that supports creativity and continual innovation is awaiting you. Collaborative ownership and freedom with responsibility is characteristic for the way we act and work as a team.

At Axelera AI, we wholeheartedly embrace equal opportunity and hold diversity in the highest regard. Our steadfast commitment is to cultivate a warm and inclusive environment that empowers and celebrates every member of our team. We welcome applicants from all backgrounds to join us in shaping the future of AI.