LLMs can't optimize schedules, but AI can!
summary
Argues LLMs fail at combinatorial planning (30 employees → 150 shifts, search space ~30^150) and shows a working alternative: metaheuristics via the Java framework Timefold that uses annotated domain classes, hard/soft constraints and local search. Live demos include a production-like run on ~1,000 shifts for 50 employees with reproducible solves and tens of thousands of moves per second, a modeling pipeline using PlantUML and agents to generate code, and a customer case where route optimization cuts driving time ~25% (≈10M kg CO2 / ~$100M saved).
AI-written summary · about
chapters (18)
- 0:00Introduction
- 2:28GenAI limitations
- 3:31Planning problems
- 8:16LLMs fail at scheduling
- 11:04Why LLMs struggle
- 14:56AI beyond ML
- 20:12Math optimization
- 21:59Introducing Timefold Solver
- 23:24Timefold Solver demo
- 27:20How Timefold works
- 31:36Use cases
- 35:31Real-world impact
- 36:58Future of AI
- 38:23LLMs for modeling
- 41:38AI code generation
- 42:21AI improving algorithms
- 43:24Cloud Code integration
- 44:04Conclusion & Q&A
related talks
41mBuilding LLM Agents – Challenges and Learnings
19mLLMs reasoning and agentic capabilities over time
28mSymbolic Reasoning in telecom oriented LLMs
26mAI Agents as Financial Co-Pilots | AI-Driven Strategy & Decision Making
29mLocal LLM fine-tuning: a practical example
13mThe 2025 AI Engineering Report
40mLLMs - application and impact on business
46mPlaying the long game