Current

Automated Agentic workflow for Scientific Discovery

Automation of agentic workflows that are system aware to automate scientific discoveries. Our framework, MARGE (built on SciAgents), jointly synthesizes workflow DAGs from natural-language scientific goals and maps each task to local or remote HPC resources based on live system conditions, modeling compute cost, data movement, communication locality and utilization. A dual-variable refinement layer keeps reshaping the workflow and its placement during optimization, and runtime introspection with cross-site parallel execution adapts to changing HPC environments. Across scientific use cases (e.g., ERA5 climate, genomics) on ALCF Polaris, Jetstream and PNNL systems, MARGE reduces workflow makespan by up to 83.78%.

Small Scale Edge LLM's

Develop efficient systems for task division including computation offloading, routing and and forwarding among edge, cloud and cluster

ML Compilers & GPU Kernel Tuning

Where do ML compiler defaults leave GPU performance on the table, and is the gap about choosing configurations or about backend code generation? We run an empirical study of IREE GEMM code generation and autotuning on NVIDIA A100 (SIMT FP32) and AMD MI300X (MFMA F32): tuned vs. default performance across a curated GEMM benchmark set, Speed-of-Light and hierarchical roofline analysis, cost models that cut tuning search cost, knob ablations, LLM-decode shapes, and a cross-version comparison with SharkTuner.

Past

LLM's for timeseries

Currently we have been working on LLM based forecasting and unsupervised clustering for time series data

AI Guided Cloud Scheduler

We are working on ML based scheduling for RAPS from ORNL

- ML for System

We are working on ML based scheduling for RAPS from ORNL

Page updated

Google Sites

Report abuse