Step away from the vibe coding, take a break from your prompt-and-pray sessions: it's time to reconnect in real life. Join us for the 26th Belgium NLP Meetup to get a pulse on what's happening in the Belgian NLP landscape. Three speakers will talk about how they tackle real-world challenges in knowledge retrieval, software security, and business automation with NLP.
We're gathering at Silverfin in Ghent on Thursday, June 12th. Doors open at 7pm with pizza and drinks to kick things off. Talks begin around 7.30pm, and from 9pm onward we welcome you to stay for informal networking and conversation.
Unhobbling the R in RAG
Laurent Sorber (Superlinear)
Retrieval Augmented Generation (RAG) is perhaps the most commercially successful application of Large Language Models (LLMs) because it enables users to get direct answers to their questions from any existing repository of knowledge. However, vanilla RAG often disappoints users with irrelevant, incorrect, or incomplete answers. A RAG system’s output can only be as good as the information that it retrieves. In this talk, we’ll identify the weaknesses of vanilla RAG, and discuss the best solutions to those problems. Together, these improvements yield a modern RAG pipeline that unhobbles the R in RAG.
LLMs in Cybersecurity
Berg Severens (Aikido Security)
Vulnerability detection in source code is typically based on hard-coded patterns. This is a powerful method for detecting a lot of ways hackers might use to attack your system, but it comes with a lot of false alerts. LLMs can understand more context of the code and can significantly reduce the number of false positives. Moreover, they can also prioritize vulnerabilities, so companies can focus on the lowest hanging fruit in reducing cybersecurity debt.
Automating account mapping with neural networks: a multiclass classification approach
Wouter Dobbels (Silverfin)
Streamlining accounting across numerous companies requires mapping their diverse charts of accounts (CoAs) to a standard one – often containing thousands of entries. This traditionally manual and tedious mapping process is crucial but slow. We’ve automated this by framing it as a multiclass classification problem, using neural networks to predict the standard CoA category based primarily on an account’s name and number. This talk will cover our journey from the initial model to overcoming significant technical challenges, including supporting multiple standard CoAs with a single model, handling multiple languages, and implementing online learning.
