First Zürich AI Meetup

Join us for first Zürich AI Meetup

Agenda:
- The risks of LLM hallucinations: Finetuning vs RAG by João Carvalho (UTHEREAL | ML PhD at ETH AI Center)

- Leveraging AI for Enhanced User Experience by Ferdinand Vogler (Ricardo)

- Prompt Mutation Pipeline – Improving Results Without Fine-Tuning the Model by Michał Koźmiński (Beekeeper)

Abstract:

The risks of LLM hallucinations: Finetuning vs RAG
One of the key challenges when deploying large language models (LLMs) in real-world applications is their tendency to generate "hallucinations"—factually inaccurate outputs based on plausible but incorrect combinations of input prompts and responses. In this presentation, we'll explore the causes of these hallucinations and examine two approaches to mitigate them: model fine-tuning and retrieval-augmented generation (RAG). We'll compare their effectiveness, discuss real-world applications, and highlight best practices for minimizing risks in LLM-based systems.

Leveraging AI for Enhanced User Experience
Practical Applications in Search, Personalization, and Fraud Prevention at Switzerland’s largest second-hand marketplace

Nearly every Swiss household has an account on the second-hand marketplaces Ricardo, Tutti or Anibis. This talk will explore how AI is changing user interactions through improved search algorithms, personalised recommendations, and robust fraud prevention.

Ferdinand will showcase real-world case studies, demonstrating AI’s role in helping users discover their favorite items, tailoring their experiences, and catching fraudsters. By separating false promises from practical applications that are used by hundreds of thousands of users every day, attendees will learn how AI can improve their platforms.

Prompt Mutation Pipeline – Improving Results Without Fine-Tuning the Model
As large language models (LLMs) continue to evolve, improving their performance without costly fine-tuning becomes crucial for scalability and efficiency. In this presentation, Michał Koźmiński will introduce the “Prompt Mutation Pipeline,” a novel approach that enhances LLM results through automated prompt evaluation and modification. By leveraging both organic and synthetic data, this method provides a cost-effective alternative to traditional fine-tuning, potentially reducing costs by up to 90%. Join this session to discover practical techniques for optimizing LLM outputs and improving performance in real-world applications without modifying the underlying model.