Smart Suggestion Framework on Optimizing Machine Learning on TPUs and GPUs

• Developed a rule-based suggestion framework in XProf to diagnose distributed model-training bottlenecks across TPUs/GPUs and provide actionable optimization recommendations.
• Built an agentic interface on Gemini that analyzes training sessions and delivers personalized guidance to improve model-training efficiency.

Packet-level AI-network Simulation Workload Support with LLM Self-Service

• Developed and deployed support for distributed training topology simulations with Meta’s internal LLM agents, enabling faster profiling of various distributed training setups.
• Designed a generalized conversion template to simulate newly proposed network topologies seamlessly in the framework.

Unified Spatial-Temporal Multi-Vehicle Collaborative Perception

• Benchmarked performance of existing collaborative perception methods under sensor localization and synchronization errors.
• Designed a transformer-based model to fuse multi-vehicle features with tolerance to data misalignment caused by GPS measurement errors, network transmission latency, and network packet loss.