Principal Software Engineer - AI Infrastructure

Microsoft • Redmond, WA

1d • $142,800 - $304,200

About The Position

Microsoft’s Cloud + AI (C+AI) organization builds the end-to-end AI stack powering innovation across Azure and Microsoft. Within the Foundry, Inference and Training (FIT) Training Team, this role helps build infrastructure for training Large Language Models (LLMs), Small Language Models (SLMs), multimodal, code-specific, agentic, proprietary, and open-weight models using open-source training frameworks.

The Principal Software Engineer owns the architecture and technical direction of major training-platform components, taking ambiguous problems through production operation. The role partners with researchers and product teams to understand their needs, reduce friction, and measure success through customer outcomes. It provides technical leadership through design, code, and test-strategy reviews, mentorship, and coordination with teams across C+AI and partner organizations. The position also contributes to large-scale distributed AI training systems, model training and evaluation, production services, and engineering practices in a fast-moving environment.

Requirements

• Bachelor’s degree in Computer Science or a related technical field and 6+ years of technical engineering experience coding in languages such as C, C++, C#, Java, JavaScript, or Python, or equivalent experience.
• 5+ years designing, building, and operating large-scale distributed systems or cloud platforms with end-to-end ownership of production services, including Kubernetes-based containerized platforms on Azure, AWS, or Google Cloud; control-plane or backend services managing compute, storage, and networking; and secure multi-tenant designs covering authentication, authorization, workload identity, quota management, and isolation.
• 3+ years building machine-learning training infrastructure, such as GPU cluster scheduling, data pipelines, and checkpointing and storage systems, using open-source training frameworks such as PyTorch FSDP, DeepSpeed, Megatron-LM, Hugging Face Transformers and Accelerate, or Ray Train.
• 2+ years training, fine-tuning, or evaluating open-weight models at multi-node scale.
• 2+ years building or operating RL post-training pipelines for LLMs and agentic models using verl, slime, MILES, or comparable frameworks, with understanding of the end-to-end training stack, including pretraining and post-training frameworks, RLEs, rollout inference engines such as vLLM or SGLang, evaluation, and checkpointing, and how design choices affect throughput, stability, and cost.
• 3+ years serving as an on-call or DRI engineer for production services with defined availability or latency targets, including incident response, postmortems, structured logging, distributed tracing, dashboards, alerting, automated validation, and incident playbooks.
• 2+ years partnering directly with internal or external customers to define requirements, instrument customer experience (such as adoption, satisfaction, or time-to-first-result metrics), and prioritize engineering work by customer impact.
• Experience leading technical design across 2+ teams, driving ambiguous multi-quarter initiatives to completion, and mentoring senior engineers.

Nice To Haves

• Master’s degree in Computer Science or a related technical field and 8+ years of technical engineering experience coding in languages such as C, C++, C#, Java, JavaScript, or Python; or a bachelor’s degree in Computer Science or a related technical field and 12+ years of such experience; or equivalent experience.
• Contributions to open-source training framework projects.

Responsibilities

• Own the architecture and technical direction of major components of the training platform for LLMs, SLMs, multimodal, code-specific, and open-weight models, ensuring scalability, reliability, security, and cost efficiency.
• Partner with researchers and product teams across Microsoft to understand training and evaluation needs, translate them into platform requirements, and measure customer experience and outcomes.
• Drive ambiguous, multi-quarter initiatives from proposal through production, establish milestones, identify dependencies, and coordinate ownership and delivery commitments across C+AI partner teams.
• Build on and extend open-source training frameworks, including PyTorch, DeepSpeed, Megatron-LM, and Hugging Face libraries, adapting them for proprietary and open-weight models and contributing improvements where appropriate.
• Design and build production services that meet Microsoft security, privacy, Responsible AI, and compliance requirements, including auditable compliance evidence.
• Debug and resolve complex issues in distributed training jobs and production services; participate in the on-call rotation as a Designated Responsible Individual (DRI), lead incident retrospectives, and improve troubleshooting guides, testing, safe deployment, and monitoring.
• Mentor engineers and raise the team’s technical bar through design and code reviews and ownership of the team’s test strategy.
• Champion AI-assisted engineering tools and practices across the software development lifecycle.

What This Job Offers

• Job Type: Full-time
• Career Level: Principal