Supervisor(s): Arseniy Kuznetsov

Abstract: 

Large Language Models (LLMs) are rapidly evolving into tools that can interact with software, files, databases, and computing environments. Increasingly capable open-weight models can now be deployed locally, keeping sensitive data and workflows within an organization's infrastructure.

This internship aims to investigate how far modern open-weight LLMs can be used on a dedicated local workstation and to assess their usefulness for typical ESO workflows. This could include analyzing engineering telemetry, investigating logs, retrieving and processing data from adaptive optics systems, generating scripts, extracting technical information from reports, and producing analysis report from telemetry or measurement data.

The candidate will deploy and evaluate several open-weight models and inference frameworks, studying the trade-offs between selected harness, model hyperparameters, and performance. Emphasis will be placed on agentic workflows. They will also develop a small set of realistic evaluation tasks and quantitatively compare models and agent configurations on reliability, performance, resource requirements, and safe tool use.

Expected experience:

• Python and Linux
• Basic knowledge of machine learning and LLMs
• Experience with containeres, databases, or LLM inference frameworks would be an advantage