Distributed AI Systems Researcher
Accelerator communication, memory movement, and collective operations.
I work on distributed AI data-plane systems: the software that moves tensors and model state across accelerator memory, GPUs, nodes, and high-speed networks.
My interests include collective communication, GPU-aware transfers, topology-aware transport, communication–computation overlap, and reproducible performance analysis.
I am a Google Summer of Code 2026 contributor with the STE||AR Group, working on hpx::collectives.
- Implemented hierarchical
all_reduce,all_gather,all_to_all, and prefix scans. - Diagnosed serialization, centralized data-path, and transport-threshold bottlenecks.
- Reduced large-message
all_to_allperformance from 7.1× behind OpenMPI to approximately 1.2×. - Added contiguous multidimensional payloads, communicator-generation management, benchmarks, and distributed regression tests.
A model-free regression canary for distributed-LLM communication that preserves configuration rankings, regression decisions, and latency-tail behaviour.
communication trace → canary → replay → verify
- Accelerator data movement and GPU-aware communication
- Collective algorithms and distributed runtime systems
- Memory registration, staging, and asynchronous transfers
- Communication–computation overlap
- Cluster-scale performance profiling
C++20 · Python · CUDA · HPX · MPI · NCCL · LCI · Triton · Linux · Slurm