Software Engineer, Systems — AI Training Data Infrastructure

Job description

We own the dataset layer that Meta's largest AI training runs read from. AIRStore and the Anywhere Training substrate are what let a multi-petabyte dataset be written once and read at full throughput from any cluster, in any region, in any cloud — without a copy. Our customers are named model programs, not abstract services: when a training job's GPUs go idle waiting on I/O, or a dataset isn't where the scheduler put the job, that is our problem to own and fix. In 2026 this team moved Anywhere Training blob pointers to 100% Manifold residency, cut AIRStore dataset startup time by 10×, drove the migration of AIRStore datasets onto a standard S3 interface, and reclaimed hundreds of petabytes through lifecycle work — all while holding the line on training reliability across dozens of production incidents.