Fall 2026 · Build what you use
Course projects in AI Infrastructure. Start with a systems challenge, implement your solution, and support every performance claim with reproducible evidence.
Explore the projects
Implement. Evaluate.
Make the result reproducible.
Build an automated optimization workflow that improves serving cost-effectiveness under TTFT and TPOT constraints, with disaggregated prefill and decode as the primary deployment.
Build a reproducible reinforcement-learning pipeline in which the model and harness evolve jointly, and compare fixed, separate, and joint optimization under comparable budgets.
Transform PyTorch subgraphs into fused forward and backward kernels, preserving numerical correctness while reducing memory traffic and improving throughput.
Design a parallelization scheme and implement a kernel that overlaps GEMM with All-to-All communication, improving end-to-end throughput over a sequential baseline.
Implement KDA chunk prefill with Q/K normalization and develop a reusable model for evaluating kernel partitioning, data movement, and scheduling.
Prepare your own data and train Qwen3.5-0.8B with SFT and RL. Evaluate multiple-choice and open-ended questions across all 14 course weeks and demonstrate an end-to-end infrastructure task.
Upload your code, correctness evidence, and report. Save a receipt to track your submission.
Follow the full project objective and the supplied test contract. Deadlines and team policies follow course announcements.