Altana builds compound GenAI systems on Databricks to classify goods for global trade compliance

Overview

Altana, a value chain management platform, combined custom deep learning models, RAG-based fine-tuned agent workflows and RLHF on the Databricks Data + AI Platform to generate tax and tariff classifications and legal justifications for cross-border goods, and to enrich its global supply chain knowledge graph. Using MLflow, Model Serving, Delta Lake and Unity Catalog, Altana now trains and deploys GenAI models 20x faster and achieved 20-50% better model performance across its most demanding workloads.

The challenge

Altana had to create boilerplate tools for launching GenAI applications, diverting focus from product functionality, and faced tension between product functionality, total cost of ownership, vendor lock-in and information security while needing to rapidly iterate on approaches and fine-tune data and models across several cloud deployments for diverse customers.

The solution

Altana used the Databricks Data + AI Platform's open architecture to combine custom deep learning models with fine-tuned agent RAG workflows and RLHF refinement, using Managed MLflow and Model Serving to manage the ML lifecycle, Delta Lake for real-time data ingestion and time-travel debugging, and Unity Catalog for federated data governance and customer data privacy.

Reported business value

Altana deployed GenAI models 20 times faster and achieved 20-50% better model performance, even for its most demanding workloads, enabling customer-centric innovations that were previously unattainable.