EBOS Group is a leading marketer and distributor of healthcare and pharmaceutical products across Australia, New Zealand, and Southeast Asia, with a $12 billion annual revenue and over 5,700 employees. A critical operational bottleneck involves processing millions of handwritten pharmacy charts and scripts annually. Accuracy is non-negotiable, yet the diversity of handwriting styles, complex form layouts, and the need for clinical-grade validation — such as matching medication entries to authorized signatures — presented a challenge that general-purpose cloud LLMs could not solve.

Initial pilots with public cloud AI failed due to:

• Insufficient accuracy
• Prohibitive financial costs at scale
• The inability to meet strict Australian data residency and sovereign compliance requirements

In this session, we reveal how EBOS and John Snow Labs engineered a Databricks-native, multimodal, production-grade pipeline designed for medical image processing. The architecture moves beyond GenAI experimentation into a mission-critical system, utilizing:

• Multimodal Healthcare AI: Leveraging John Snow Labs’s Medical VLLM, Visual NLP, and Healthcare NLP models to extract high-fidelity data from complex, multi-page medical documents.
• Deterministic Reliability: Unlike probabilistic general LLMs, the pipeline utilizes deterministic NLP and Small Language Models (SLMs). This ensures reliable normalization of clinical entities and medication names, which is essential for reducing drift in high-risk pharmaceutical workflows.
• Sovereign Infrastructure & Serving: To ensure data residency, the solution is deployed within the Databricks AU region. We utilize Databricks Model Serving on A100 GPUs — a strategic choice to ensure consistent resource availability and lower operational costs compared to the scarce H100 clusters in the region.
• End-to-End Governance: The entire lifecycle is governed via Unity Catalog, providing the lineage tracking, auditability, and access controls required for sensitive patient information.

Attendees will gain a practical blueprint for operationalizing multimodal AI in regulated environments. We will share lessons on:

• Optimizing A100-based serving for clinical latency
• Balancing CPU vs. GPU compute to manage ROI
• Building a quality assurance bridge between raw model capability and real-world clinical deployment