Providence serves vulnerable and disadvantaged communities through compassionate, high-quality care. As one of the largest nonprofit health systems in the United States—with 51 hospitals, over 1,000 outpatient clinics, and more than 130,000 caregivers across seven states—our ability to deliver timely, coordinated care depends on transforming not only clinical outcomes but also the workflows that support them.
One of the most pressing instances is automating the way we handle faxes. Despite advances in digital health, faxes remain a dominant form of communication in healthcare, especially for referrals between providers. Providence receives more than 40 million faxes annually, totaling over 160 million pages. A significant portion of that volume must be manually reviewed and transcribed into Epic, our electronic health record (EHR) system.
The process is slow, error-prone and contributes to multi-month backlogs that ultimately delay care for patients. We knew there had to be a better way.
Tackling messy workflows and unstructured data at scale
The core challenge wasn’t just technical—it was human. In healthcare, workflows vary widely between clinics, roles and even individuals. One staff member might print and scan referrals before manually entering them into Epic, while another might work within an entirely digital queue. The lack of standardization makes it difficult to define a “universal” automation pipeline or create test scenarios that reflect real-world complexity.
On top of that, the underlying data is often fragmented and inconsistently stored. From handwritten notes to typed PDFs, the diversity of incoming fax documents creates a wide range of inputs to process, classify and extract information from. And when you’re dealing with multiple optical character recognition (OCR) tools, prompt strategies and language models, tuning all these hyperparameters becomes exponentially harder.
This complexity made it clear that our success would hinge on building a low-friction testing ecosystem. One that lets us experiment rapidly, compare results across thousands of permutations and continuously refine our models and prompts.
Accelerating GenAI experimentation with MLflow on Databricks
To meet that challenge, we turned to the Databricks Data Intelligence Platform, and specifically MLflow, to orchestrate and scale our machine learning model experimentation pipeline. While our production infrastructure is built on microservices, the experimentation and validation phases are powered by Databricks, which is where much of the value lies.
For our eFax project, we used MLflow to:
- Define and execute parameterized jobs that sweep across combinations of OCR models, prompt templates and other hyperparameters. By allowing users to provide dynamic inputs at runtime, parameterized jobs make tasks more flexible and reusable. We manage jobs through our CI/CD pipelines, producing YAML files to configure large tests efficiently and repeatably.
- Track and log experiment results centrally for efficient comparison. This gives our team clear visibility into what’s working and what needs tuning, without duplicating effort. The central logging also supports deeper evaluation of model behavior across document types and referral scenarios.
- Leverage historical data to simulate downstream outcomes and refine our models before pushing to production. Catching issues early in the testing cycle reduces risk and accelerates deployment. This is particularly important given the diversity of referral forms and the need for compliance within heavily regulated EHR environments like Epic.
This process was inspired by our success working with Databricks on our deep learning frameworks. We’ve since adapted and expanded it for our eFax work and large language model (LLM) experimentation.
While we use Azure AI Document Intelligence for OCR and OpenAI’s GPT-4.0 models for extraction, the real engineering accelerant has been the ability to run controlled, repeated tests through MLflow pipelines—automating what would otherwise be manual, fragmented development. With the unifying nature of the Databricks Data Intelligence Platform, we’re able to transform raw faxes, experiment with different AI techniques and validate outputs with speed and confidence in one place.
All extracted referral data must be integrated into Epic, which requires seamless data formatting, validation and secure delivery. Databricks plays a critical role in pre-processing and normalizing this information before handoff to our EHR system.
We also rely on Databricks for batch ETL, metadata storage and downstream analysis. Our broader tech stack includes Azure Kubernetes Service (AKS) for containerized deployment, Azure Search to support retrieval-augmented generation (RAG) workflows and Postgres for structured storage. For future phases, we’re actively exploring Mosaic AI for RAG and Model Serving to enhance the accuracy, scalability and responsiveness of our AI solutions. With Model Serving, we will be in a better position to effectively deploy and manage models in real time, ensuring more consistent workflows across all our AI efforts.
From months of backlog to real-time triage
Ultimately, the beneficiaries of this eFax solution are our caregivers—clinicians, medical records administrators, nurses, and other frontline staff whose time is currently consumed by repetitive document processing. By removing low-value manual bottlenecks, we aim to return that time to patient care.
In some regions, faxes have sat in queues for up to two to three months without being reviewed—delays that can severely impact patient care. With AI-powered automation, we’re moving toward real-time processing of over 40 million faxes annually, eliminating bottlenecks and enabling faster referral intake. This shift has not only improved productivity and reduced operational overhead but also accelerated treatment timelines, enhanced patient outcomes, and freed up clinical staff to focus on higher-value care delivery. By modernizing a historically manual workflow, we’re unlocking system-wide efficiencies that scale across our 1,000+ outpatient clinics, supporting our mission to provide timely, coordinated care at scale.
Thanks to MLflow, we’re not just experimenting. We’re operationalizing AI in a way that’s aligned with our mission, our workflows, and the real-time needs of our caregivers and patients.
