A prototype demonstrating how KurrentDB's event sourcing architecture can serve as training data for fine-tuning small language models on domain-specific business processes. Using Claude Code skills (KurrentDB and Hugging Face Model Trainer), the author fine-tuned SmolLM2-360M-Instruct with LoRA on 1,487 financial process event sequences in roughly one hour. The resulting model was deployed as an agentic anomaly detector that achieved 100% recall — catching all 50 injected process anomalies — with only 27 minutes of training on consumer GPU hardware. The article covers event schema design, training data generation from event streams, LoRA fine-tuning with TRL/SFTTrainer, and an agentic orchestrator that monitors live event streams and flags unexpected transitions with explanations.
