AI, LLM, and Intelligent Agents: The Future of Integration with Quarkus

Artificial intelligence is revolutionizing the software development landscape, but how reliable are Large Language Models (LLMs)? This question was at the heart of the event "Integrating LLMs and Intelligent Agents: AI with Quarkus", organized by the Java User Group Torino in collaboration with Zero11. The event provided deep technical insights into combining generative AI, rule engines, and validation tools to build more reliable and efficient systems.

Two outstanding speakers led the discussion:

• Mario Fusco, Java Champion, Senior Principal Software Engineer @Red Hat, Project Lead of Drools
• Daniele Zonca, AI Architect @Red Hat OpenShift AI, Co-founder of TrustyAI, contributor to KServe, vLLM, and Kubeflow

LLMs: Powerful but Not Infallible

Mario Fusco kicked off the discussion by emphasizing that LLMs are transforming software development but have inherent limitations. Their operation is based on statistical algorithms that, while generating coherent responses, lack an understanding of business logic and operational context.

"If you buy a washing machine on Amazon, the next day the system suggests buying three more. This makes sense statistically, but not for a regular customer who isn’t opening a laundromat," Fusco humorously pointed out.

This highlights a crucial issue: without a structured business logic layer, LLMs can generate recommendations or decisions that don’t align with real-world business needs.

The Problem of AI Hallucinations

One of the central themes of the event was hallucinations in AI models, where LLMs produce incorrect or unverifiable responses. Daniele Zonca elaborated on this challenge, citing an infamous case involving Air Canada.

"A chatbot for the airline promised a customer a $650 refund, even though company policy didn’t allow it. The customer took the transcript to court, and Air Canada was forced to pay. The next day, the chatbot was shut down."

This case underscores why blindly trusting LLMs without structured oversight can lead to disastrous consequences.

Integrating Symbolic AI and Machine Learning: The Hybrid Approach

To mitigate these risks, the most promising approach is a hybrid model that combines generative AI with rule engines and other validation tools.

"We already have all the necessary technologies: rule engines have been around for over 40 years, and now we can integrate them with LLMs to get the best of both worlds," Fusco explained.

The key idea is to pair statistical generation with rigid rule-based validation, ensuring that AI-driven decisions are transparent and consistent. Instead of letting LLMs determine critical business rules—such as refunds in the Air Canada case—companies should use rule engines like Drools to enforce business logic and prevent incorrect outputs.

Live Demo: AI-Powered Refund Processing Chatbot

One of the event’s highlights was a live demo showcasing a refund management chatbot built with Quarkus, LangChain4j, and Drools.

The system collected user input via chat, then passed the data to Drools, which applied business rules to accurately calculate the refund amount.

"The LLM is used only for data collection and user interaction, while all business logic is handled by the rule engine," Fusco explained.

The Power of Quarkus and LangChain4j

The event emphasized the critical role of Quarkus and LangChain4j in building scalable and reliable AI applications. These tools allow developers to integrate LLMs with rule-based validation systems, improving control over AI-generated outputs.

"Quarkus is ideal for managing hybrid AI architectures, while LangChain4j provides powerful tools for orchestrating interactions between generative models and intelligent agents," Zonca pointed out.

Key Takeaways: Towards a More Reliable AI

The event confirmed that the future of AI won't be solely based on generative models but on a synergy between LLMs, rule engines, and advanced validation techniques.

"AI should not be seen as a monolithic system. True value comes from integrating different technologies, each with its strengths," Fusco concluded.

Zero11 is proud to have supported this event and actively contributes to the development of cutting-edge AI solutions. Our mission is to develop and integrate advanced technologies to enhance sales and marketing processes while ensuring transparency, efficiency, and control over AI systems.

📌 Missed the event? Watch the full session here.

🔗 Want to explore how to integrate generative AI and rule engines into your business? Contact us to learn more about Zero11’s AI-powered solutions!

#AI #MachineLearning #LLM #Quarkus #Drools #LangChain4j #Zero11 #ArtificialIntelligence #RuleEngine #Developers #Innovation