AI for HPC Operations (2024 - Present)

Overview

Bringing AI into HPC operations requires translating unstructured machine data into structured insights. This project explores how LLMs can automate data science workflows, reducing the dependency on domain-specific expertise.

• Built an LLM-based query system allowing operators to interact with HPC telemetry using natural language
• Designed a self-adaptive AutoML system for operational forecasting, reducing the barrier for predictive analytics in HPC centers.
• Established an autonomous model registry and retraining pipeline for continuously evolving operational intelligence.

Technology

• Python frameworks: Langchain, Chainlit, pydantic, ChromaDB, LiteLLM, HuggingFace
• Retrieval Augmented Generation ingest pipeline and context
• Pytorch based predictive ANN
• AutoML frameworks (AutoGluon Tabular and TimeSeries)
• A combination local LLM models such as Llama 3.1 and 3.2 and large OpenAI models
• Context aware tool call driven hierarchical LLM chains structure
• DuckDB based SQL interface to parquet based telemetry data
• MLflow based experiment and model tracking

Publications

• Publication preparation targeting SC25 (Work in progress)