Member of Technical Staff - Senior ML Engineer - MAI Super Intelligence Tea at Microsoft

Zurich, Zurich, Switzerland -

Full Time

Start Date

Immediate

Expiry Date

15 May, 26

Salary

0.0

Posted On

14 Feb, 26

Experience

2 year(s) or above

Remote Job

Yes

Telecommute

Yes

Sponsor Visa

Skills

Vision-Language Models, VLM Research, Inference Optimization, Kernel Optimization, CUDA, HIP, SGLang, vLLM, Knowledge Distillation, Pruning, Quantization, NVIDIA H100, AMD MI300x, Multilingual Alignment, Document Understanding, Visual Question Answering

Industry

Software Development

Description

Overview We are seeking a Senior Machine Learning Engineer to bridge the gap between advanced Vision-Language Model (VLM) research and high-performance production serving. Unlike standard data science and engineering roles, this position requires a dual competency: you must be capable of designing novel VLM architectures (including dataset curation and multilingual alignment) AND optimizing the inference stack (kernel optimization, distillation, and memory management) to run these models on specific hardware constraints (NVIDIA H100 and AMD MI300x). The successful candidate will own the entire vertical slice: from reading the latest arXiv papers and improving training sets, to writing the C++/CUDA kernels that serve the final model in production. Responsibilities 1. VLM Research & Architecture Design Continuously evaluate and implement the latest research trends in Vision-Language Models, specifically focusing on Referring Expression Comprehension (REC), Document Understanding (Pix2Struct), and Visual Question Answering (VQA).Design and build massive-scale training and evaluation datasets, ensuring multilingual compatibility and broad visual understanding for European market requirements.Lead the model co-design process, creating architectures that are natively optimized for accelerator capabilities (compute-bound vs. memory-bound operations). 2. Advanced Inference Optimization & Serving Architect high-throughput serving layers using SGLang and vLLM, optimizing for non-standard decoding strategies. Implement scientific experiments to find the Pareto-optimal frontier between serving latency and generation quality.Execute Knowledge Distillation (KD), unstructured pruning, and quantization techniques to fit large-scale VLM architectures onto single-node GPU setups (specifically H100 or MI300x) without compromising model quality. 3. Systems Engineering & Kernel Development Write and optimize custom kernels (CUDA/HIP) to accelerate serving latency, identifying bottlenecks at the operator level. Manage the full pre-training and post-training tech stack, ensuring seamless integration between model weights and inference engines.Take ownership of landing the serving-efficient model in a production environment, ensuring reliability and scalability. Qualifications Mandatory Requirements (Must Have) Education: Master’s or PhD in Computer Science, Artificial Intelligence, or High-Performance Computing. Experience: Minimum 4+ years of experience in Machine Learning, with a mandatory split focus between Model Architecture and Systems Optimization. VLM Expertise: Proven experience building and shipping Vision-Language Models (e.g., architectures similar to CLIP, Flamingo, Pix2Struct). Must have experience creating custom evaluation sets for tasks like Document Understanding. Serving Stack Proficiency: Expert-level knowledge of SGLang and vLLM for optimized serving. Hardware Specifics: Demonstrable experience optimizing models for both NVIDIA (H100) and AMD (MI300x) accelerators. Optimization Techniques: Hands-on experience with Knowledge Distillation and Pruning to reduce model latency for target serving sizes. Production Engineering: A track record of taking complex multi-modal models from research code to a deployed, user-facing production product. This position will be open for a minimum of 5 days, with applications accepted on an ongoing basis until the position is filled. Microsoft is an equal opportunity employer. All qualified applicants will receive consideration for employment without regard to age, ancestry, citizenship, color, family or medical care leave, gender identity or expression, genetic information, immigration status, marital status, medical condition, national origin, physical or mental disability, political affiliation, protected veteran or military status, race, ethnicity, religion, sex (including pregnancy), sexual orientation, or any other characteristic protected by applicable local laws, regulations and ordinances. If you need assistance with religious accommodations and/or a reasonable accommodation due to a disability during the application process, read more about requesting accommodations.

Responsibilities

The role involves bridging VLM research and production serving by designing novel architectures, curating multilingual datasets, and optimizing the inference stack for specific hardware like H100 and MI300x. Responsibilities span from research implementation and dataset creation to writing custom C++/CUDA kernels for production deployment.