Software Engineer, Systems — AI Training Data Infrastructure

Own significant components of the AIRStore data path end to end — ingestion, metadata, client, and read path — from design through production operation Attack throughput and latency as a first-class product concern: prefetching, parallelism, caching, and startup cost, measured in GPU utilization and training wall-clock rather than microbenchmarks Build the multi-region and multi-cloud story: make dataset location invisible to the training job, across Meta data centers and third-party clouds Get dataset lifecycle right — TTL, archival, expiration, and deletion — where the correctness bar is absolute in both directions: nothing a live run needs may disappear, and nothing that must be deleted may persist Own a widely embedded client library responsibly: compatibility, rollout safety, and blast-radius control across thousands of callers you don't control Take real operational ownership.