Sai Teja Gurrapu

Scaled Meta’s GPU infrastructure powering LLMs (e.g., LLaMA) and enabled efficient distributed training, boosting AI research productivity and reducing model development cycles. Built high-throughput pipelines across 24,000+ H100 GPUs and contributed to global AI hardware standards that cut training costs by 20%. Implemented distributed training workflows in PyTorch with tensor and pipeline parallelism, and developed AI-assisted data pipelines for preprocessing, fault recovery, and experiment tracking around billions of parameters. Designed monitoring with AutoGen agents for intelligent workload management and dynamic scaling, and created custom schedulers coordinating thousands of nodes for optimized resource allocation.

AI/ML Engineer responsible for scaling Meta’s GPU infrastructure powering LLMs such as LLaMA and enabling efficient distributed training. Built high-throughput compute pipelines across 24,000+ H100 GPUs, accelerating AI-assisted pipelines and production deployments for global research teams. Contributed to open AI hardware standards, reducing training costs by ~20% and advancing scalability. Engineered PyTorch distributed training workflows with tensor and pipeline parallelism to optimize GPU utilization during pretraining, fine-tuning, and large-scale inference. Developed AI-assisted pipelines automating dataset preprocessing, fault recovery, and experiment tracking for billions of parameters. Designed GPU-cluster monitoring with AutoGen agents for workload management and anomaly detection, enabling dynamic scaling. Implemented custom schedulers coordinating thousands of nodes and integrated PyTorch Distributed with AI tools. Collaborated with research teams on next-generation LLaMA

Scaled Meta’s GPU infrastructure powering LLMs like LLaMA; delivered high-throughput compute pipelines across 24,000+ H100 GPUs; contributed to global AI hardware standards; engineered distributed training workflows in PyTorch with tensor and pipeline parallelism; developed AI-assisted pipelines automating dataset preprocessing, fault recovery, and experiment tracking; designed and deployed GPU cluster monitoring with AutoGen agents; implemented custom schedulers coordinating thousands of nodes; partnered with research teams on LLaMA training/inference; built Python automation scripts for multimodal dataset orchestration and reproducible experiments across Meta’s infrastructure platforms; applied AutoGen multi-agent orchestration to scale LLaMA training and inference.