This event is organized in collaboration with Eraneos.

RSVP

RSVP on meetup.com!

Save the Date!

January 8th, 2025 6:30pm CEST – 9 pm CEST

Location

Klenzestr. 41, 80469 München - Google Maps Link

About this Event

From Scratch to Production: LLM Based E2E Service in Swiss Banking

Speakers: Jakob Klement, David Satomi, Fabian Stermann

Handling over 60 thousand customer interactions per week is hard and costly. Especially in Switzerland, where four different languages meet a high consumer expectation for quality in every interaction, surrounded by strict regulations. In this talk, we give detailed tech insights regarding the development and implementation of an end-to-end and LLM-based voice and chatbot that replaced a legacy IVR system at a swiss bank. Over the course of 1.5 years, we went from scratch to production, while facing technical and regulatory challenges, and ended up increasing customer satisfaction, as well as the first-solution rate from around 40% to more than 85%.

Beyond Retrieval: The Technical Core of Reliable Agent Reasoning

Speakers: Ludovic Tessier

This presentation moves beyond simple Retrieval-Augmented Generation (RAG) to explore what is required for the next step in AI development: reliable agent reasoning. The talk argues that while retrieving information is a solved problem, the real challenge lies in an agent’s ability to intelligently process that information. The session will provide a deep dive into the technical core components that enable an AI agent to analyze, synthesize, evaluate, and act upon retrieved data.

Advances in Open-Ended Text Generation and Evaluation for Large Language Models

Speaker: Dr. Matthias Aßenmacher

This talk explores the evolving landscape of open-ended text generation and decoding strategies for LLMs. We begin with early work on model-based evaluation metrics, which exposed critical shortcomings in their sensitivity to semantic deterioration and structural variation. Moving on, we explore multicriteria evaluation frameworks that move beyond isolated metrics by ranking decoding methods holistically using partial orders and composite cardinal metrics. Subsequently, we introduce Adaptive Contrastive Search (ACS), an uncertainty-aware decoding strategy, to improve coherence and creativity without sacrificing diversity, followed by a large-scale analysis revealing how decoding hyperparameters affect output quality across domains and tasks. Finally, we touch upon very recent work: GUARD introduces a hyperparameter-free, uncertainty-driven decoding strategy building upon ACS.