Staff Software Engineer, AI Runtime

P-1930 At Databricks, we are passionate about enabling data teams to solve the world's toughest problems — from making the next mode of transportation a reality to accelerating the development of medical breakthroughs.

We do this by building and running the world's best data and AI infrastructure platform so our customers can use deep data insights to improve their business.

Training and customizing state-of-the-art AI models is one of the most demanding workloads in computing, and it sits at the heart of Databricks' Mosaic AI mission.

AI Runtime (AIR) is our managed platform for large-scale GPU training and fine-tuning.

It gives customers on-demand access to fleets of the latest accelerators and a serverless experience that hides the complexity of provisioning, scheduling, and orchestrating multi-node jobs, with the resilience to keep training running for days or weeks across thousands of GPUs.

AIR powers the full spectrum of custom training, from fine-tuning open models to pre-training frontier-scale foundation models, for some of the most sophisticated AI teams in the world.

As a Staff Software Engineer for AI Runtime, you will play a critical role in building and scaling the systems that make large-scale training fast, reliable, and effortless.

You will drive the architecture and evolution of the managed GPU training stack, spanning scheduling and capacity, distributed training performance, fault tolerance, and the developer experience of launching and operating jobs at scale.

Beyond hands-on contributions to core systems, you will help define the long-term technical vision for AIR, mentor senior engineers, partner across product, research, and platform teams, and lead the initiatives that expand the technical and business impact of custom training at Databricks.

The impact you will have:

• Drive the architecture and evolution of AIR's managed GPU training platform, delivering scalable, high-throughput, and resilient training across fleets that span thousands of accelerators.
• Solve the hardest problems in large-scale training, including multi-node orchestration, distributed parallelism strategies, GPU scheduling and dynamic routing, high-throughput data loading, and checkpoint and restore for very long-running jobs.
• Push GPU efficiency and training performance, raising utilization (such as model FLOPs utilization and end-to-end throughput) and lowering cost per training run across diverse model architectures and hardware generations.
• Build the resilience and observability foundations that keep multi-node jobs healthy, detecting and recovering from hardware and software failures with minimal disruption to customers.