GAIA - October Meetup

Reliability in LLM Systems: Approaches for High-Stakes Applications

Welcome to this Eghed-hosted GAIA meetup!
Language models are not reliable by default: they hallucinate, and it is often hard to tell why they produced a certain answer. In knowledge-intensive and high-risk domains, where answers are costly to verify and errors have real consequences, this is a serious obstacle to deployment.

This talk looks at what it takes to build chatbots that meet strict requirements for reliability and transparency. Lovisa Hagström will go through the main techniques currently used for this - retrieval-augmented generation, context engineering, and agentic workflows - and why system architecture and rigorous evaluation deserve as much attention as the techniques themselves.

To ground this, Lovisa will draw on recent applied projects in manufacturing and document-heavy office workflows, covering the design decisions that had strong effects on reliability in practice

Speaker
Lovisa Hagström is a machine learning engineer specialising in natural language processing, currently at eghed. She holds a PhD in Computer Science from Chalmers, where her research focused on developing and evaluating reliable models for fact-critical tasks, covering language model controllability, interpretability, and prediction provenance, with publications at ACL, EMNLP, and other top NLP venues. She also spent time as a visiting researcher in the CopeNLU group at the University of Copenhagen. Her current work applies LLMs to industry-specific problems, including the projects discussed in this talk.

Agenda
17:30: Doors open & welcome
18:00: Talk by Lovisa Hagström
18:30: Food & mingle
20:00: Wrap-up
Come join us for an evening of practical insights, discussion, food, and mingling with others interested in AI and reliable LLM systems.