Fine-tune AI models with your agency data. Custom LoRA adapters with traffic routing and A/B testing for continuously improving AI performance.
XeroFlow extracts training data from your agency's actual usage — chat conversations, intent classifications, knowledge base entries, and AI interaction patterns. Five extractors produce different dataset types: QA pairs, intent examples, RAG chunks, knowledge entries, and combined sets. Data is automatically anonymized to remove PII before being used for training. Datasets export as JSONL and archive to R2.
Upload, activate, and retire LoRA adapters through the admin UI. Each adapter has metadata — training dataset, base model, creation date, and performance metrics. The system supports multiple active adapters simultaneously with weighted traffic routing, so you can gradually shift traffic from the base model to a new adapter as confidence in its performance grows.
Route a percentage of AI requests to different adapters to compare performance. Start a new adapter at 10% traffic, monitor quality metrics, and gradually increase to 100% as results validate. If an adapter underperforms, roll back instantly by adjusting traffic weights. This approach eliminates the risk of deploying a poorly trained adapter to your entire team.
Track adapter performance with metrics collected from every AI interaction — response helpfulness ratings, intent classification accuracy, response latency, and user feedback scores. Compare adapter performance against the base model and against each other. The metrics dashboard shows trends over time so you can see whether your latest training run actually improved the model for your agency's specific use cases.