CAIN 2025
Sun 27 - Mon 28 April 2025 Ottawa, Ontario, Canada
co-located with ICSE 2025
Sun 27 Apr 2025 11:40 - 11:55 at 208 - Engineering AI systems with LLMs Chair(s): Justus Bogner

Abstract—[Problem] As industries increasingly depend on complex software systems, efficient log analysis is essential for maintaining reliability and privacy. However, Identifying problems through logs is often time-consuming and costly for developers. [Background] Large language models (LLMs) can automate parts of log analysis, but challenges like limited computational resources and the frequent need to retrain LLMs due to the dynamic nature of software logs persist. External LLMs, such as GPTs, along with in-context learning techniques, can help reduce some of these issues, but other challenges, including token limitations, high token costs, and data privacy, remain. [Method] To tackle these challenges, we developed an automated pipeline that extracts log files and employs in-context learning, allowing the model to efficiently adapt to changes without extensive retraining. Our approach introduces a novel flame-graph-like method that reduces token usage, thereby lowering token-related costs and response latency while maintaining high accuracy. [Results] This solution allows industries to automate log analysis, minimize system downtime, and enhance performance, all while keeping data privacy and maintaining operational efficiency. [Conclusion] Our flame-graph-like methodology reduces input tokens by 93.61% and processing latency by 77.45%. Our anonymization results show an improvement of 138.63% over the baseline. This industrial experience report presents our approach to allow industries to balance token costs, maintain response accuracy, and ensure data privacy while relying on external LLMs without the need to manage computational resources directly.

Sun 27 Apr

Displayed time zone: Eastern Time (US & Canada) change

11:00 - 12:30
Engineering AI systems with LLMsResearch and Experience Papers at 208
Chair(s): Justus Bogner Vrije Universiteit Amsterdam
11:00
15m
Talk
Rule-Based Assessment of Reinforcement Learning Practices Using Large Language ModelsDistinguished paper Award Candidate
Research and Experience Papers
Evangelos Ntentos University of Vienna, Stephen John Warnett University of Vienna, Uwe Zdun University of Vienna
11:15
10m
Talk
Designing and implementing LLM guardrails components in production environments
Research and Experience Papers
11:25
15m
Talk
Themes of Building LLM-based Applications for Production: A Practitioner's View
Research and Experience Papers
Alina Mailach Leipzig University, Sebastian Simon Leipzig University, Johannes Dorn Leipzig University, Norbert Siegmund Leipzig University
Pre-print
11:40
15m
Talk
InsightAI: Root Cause Analysis in Large Hierarchical Log Files with Private Data Using Large Language Models
Research and Experience Papers
Maryam Ekhlasi Polytechnique Montreal, Anurag Prakash Ciena, Michel Dagenais Polytechnique Montréal, Maxime Lamothe Polytechnique Montreal
11:55
10m
Talk
Developing Multi-Agent LLM Applications through Continuous Human-LLM Co-Programming
Research and Experience Papers
Hui Song SINTEF Digital, Arda Goknil SINTEF Digital, Xiaojun Jiang Oslo University Hospital, Espen Melum Oslo University Hospital, Hyunwhan Joe Seoul National University, Caterina Gazzotti University of Modena, Valerio Frascolla Intel, Adela Nedisan Videsjorden SINTEF, Phu Nguyen SINTEF
12:05
25m
Other
Discussion
Research and Experience Papers

:
:
:
: