Experience

Research on agentic AI serving systems: program-aware inference runtimes that treat a multi-turn agent trace, rather than an isolated request, as the unit of scheduling. Building schedulers that pause, resume, and time-share long-horizon agent workloads across a pool of models on shared GPUs, with KV-cache management policies that decide per trace whether to offload conversational state to CPU memory or recompute it on resume. Investigating online routing over heterogeneous model pools (contextual bandits with delayed accuracy-latency rewards) and the architectural bottlenecks that dominate end-to-end agent latency: cache hierarchies, host-device memory traffic, and CPU-side runtime overheads.

Building end-to-end AI-driven education platforms at institutional scale, spanning full-stack development spanning data ingestion pipelines, LLM model integration, and scalable backend deployment. Designing intelligent recommendation workflows for student-advisor matching and academic analytics, incorporating retrieval-augmented generation and embedding-based similarity search. Delivering production-ready AI services with robust REST APIs serving large university-wide user bases.

Architected SkipPar, a hybrid CPU-GPU LLM training framework (IPDPSW 2026, First Author). Designed a co-execution paradigm where the GPU handles all forward and backward passes while the CPU concurrently runs parameter updates. Implemented a 4-thread producer-consumer pipeline coordinating GPU computation with CPU optimizer steps; integrated PyTorch DDP hooks for multi-GPU synchronization. Benchmarked on NVIDIA A100 & H100 GPUs using LLaMA-2 (10B) and GPT-2 (9B); achieved up to 17% reduction in end-to-end training time against AAAI- and ICPP-published baselines.

Conducted lab sessions and guided student projects on network design, backpropagation, and training dynamics.

Promoted heterogeneous programming models and parallel computing tools across the university developer community.