#VoxxedDaysCERN26

Large language models are touted as the solution for every problem under the sun. Yet, when it comes to everyday tasks like creating work schedules for 30 employees across 3 shifts or optimizing truck routes for 100 deliveries these models often derail.

We'll explore the reasons why LLMs aren't the right tool for these tasks and introduce a hidden gem when it comes to tackling these problems: metaheuristics. With live demos, you’ll see how by simply annotating your domain classes and defining your constraints, you can harness their power using Timefold, an Open Source framework built in Java.

By the end of this session, you’ll walk away with a powerful addition to your problem solver toolkit and we’ll also explore how LLMs and metaheuristics can complement each other so you can create powerful AI solutions to everyday problems.

Devoxx

LLMs struggle with combinatorial planning problems like nurse scheduling because the search space is astronomically large (30^150 possible assignments) and cannot be solved by next-token prediction or sequential reasoning. Mathematical optimization, specifically metaheuristics via the Timefold Solver framework, is the right tool for these problems. A live demo shows how to model a nurse scheduling problem in plain Java with annotated domain classes and constraint definitions (hard and soft constraints), letting Timefold explore 20,000+ moves per second to find optimal schedules. Real-world impact is illustrated with a customer achieving 25% reduction in driving time, saving $100M and 10M kg of CO2 annually. The talk also covers integrating LLMs with Timefold via agents (LangChain4j), PlantUML-based code generation, and AlphaEvolve-style code improvement loops, arguing that GenAI and classical AI techniques are most powerful when combined.

LLMs can't optimize schedules, but AI can! by Tom Cools