Research Scientist Intern, AI Infra (PhD)

About the job

Meta is seeking AI Research interns to join our Video Recommendations team within Facebook, focusing on the training and inference infrastructure behind video recommendations. The team helps people discover engaging, high-quality content across Reels, long videos, and Live streams. In this role, you will pursue a scoped research project on a selected workload or performance bottleneck, researching and building AI agents that analyze, optimize, and iteratively improve recommendation infrastructure. You will work at the intersection of machine learning, systems, and GPU computing, collaborating with Research Scientists and Software Engineers to turn research into measurable gains in training efficiency and inference performance. Our internships are twelve (12) to twenty-four (24) weeks long and we have various start dates throughout the year.

Responsibilities

Research and develop AI agents that optimize training and inference for video recommendation models, including distributed execution, data pipelines, model compilation, and serving. Investigate recursive self-improvement for video recommendation infrastructure: use profiling traces and validated experiment results to improve infrastructure performance and the agents' subsequent optimization decisions. Build GPU kernel agents that generate and tune CUDA or Triton kernels, explore memory access and kernel fusion, and verify numerical correctness and performance. Develop agentic trace analyzers that correlate CPU/GPU timelines, distributed traces, and runtime telemetry to diagnose bottlenecks and explain root causes in training and inference. Build agentic profilers for both training and inference that select profiling tools, collect targeted measurements, and recommend optimizations for compute, memory, communication, and scheduling. Design reproducible benchmarks and evaluation methods for optimization agents, measuring training throughput, inference latency and throughput, GPU utilization, memory use, and cost while preserving model quality. Collaborate with Research Scientists, Software Engineers, and infrastructure partners to scope experiments, build research prototypes, and translate validated findings into practical improvements. Communicate research plans, progress, and results through technical reports, reusable tools, and publications where appropriate.