San Francisco-based Empromptu AI launched Alchemy Models on Thursday, a platform that enables enterprises to train custom AI models using data generated from their existing production workflows. The system captures and validates outputs from enterprise applications to continuously improve AI models without requiring dedicated machine learning teams.

The platform operates by routing validated outputs from subject matter experts back into a fine-tuning pipeline that improves models over time. Companies retain ownership of the resulting model weights, which can be exported for a fee. The system differs from retrieval-augmented generation (RAG) and traditional fine-tuning by using the enterprise application itself as the continuous data source.

According to CEO Shanea Leven, the platform addresses three constraints companies face with foundation model APIs: scaling inference costs, lack of model ownership, and limited customization for domain-specific tasks. The system generates what Empromptu calls Expert Nano Models - small, task-specific models optimized for particular workflows rather than general-purpose reasoning.

Behavioral health company Ascent Autism reported significant time savings using the platform for session documentation. The company reduced documentation time from one to two hours per session to 10-15 minutes, representing up to an 87% reduction. Co-founder Faraz Fadavi cited cost reduction and improved output quality as key benefits.

The platform targets regulated and data-intensive sectors including healthcare, financial services, legal technology, and retail. Early deployments run on base models while applications accumulate sufficient production data to trigger effective fine-tuning runs. The system supports multiple base models including Llama and Qwen, with evaluation controls and compliance measures integrated into the training pipeline.