Windfall serves susceptible and deprived communities by means of compassionate, high-quality care. As one of many largest nonprofit well being programs in the US—with 51 hospitals, over 1,000 outpatient clinics, and greater than 130,000 caregivers throughout seven states—our potential to ship well timed, coordinated care is dependent upon reworking not solely scientific outcomes but additionally the workflows that help them.
Probably the most urgent cases is automating the way in which we deal with faxes. Regardless of advances in digital well being, faxes stay a dominant type of communication in healthcare, particularly for referrals between suppliers. Windfall receives greater than 40 million faxes yearly, totaling over 160 million pages. A good portion of that quantity should be manually reviewed and transcribed into Epic, our digital well being document (EHR) system.
The method is gradual, error-prone and contributes to multi-month backlogs that in the end delay take care of sufferers. We knew there needed to be a greater approach.
Tackling messy workflows and unstructured information at scale
The core problem wasn’t simply technical—it was human. In healthcare, workflows range extensively between clinics, roles and even people. One workers member would possibly print and scan referrals earlier than manually coming into them into Epic, whereas one other would possibly work inside a completely digital queue. The shortage of standardization makes it troublesome to outline a “common” automation pipeline or create check situations that replicate real-world complexity.
On high of that, the underlying information is commonly fragmented and inconsistently saved. From handwritten notes to typed PDFs, the range of incoming fax paperwork creates a variety of inputs to course of, classify and extract info from. And if you’re coping with a number of optical character recognition (OCR) instruments, immediate methods and language fashions, tuning all these hyperparameters turns into exponentially more durable.
This complexity made it clear that our success would hinge on constructing a low-friction testing ecosystem. One which lets us experiment quickly, evaluate outcomes throughout hundreds of permutations and constantly refine our fashions and prompts.
Accelerating GenAI experimentation with MLflow on Databricks
To fulfill that problem, we turned to the Databricks Information Intelligence Platform, and particularly MLflow, to orchestrate and scale our machine studying mannequin experimentation pipeline. Whereas our manufacturing infrastructure is constructed on microservices, the experimentation and validation phases are powered by Databricks, which is the place a lot of the worth lies.
For our eFax undertaking, we used MLflow to:
- Outline and execute parameterized jobs that sweep throughout mixtures of OCR fashions, immediate templates and different hyperparameters. By permitting customers to supply dynamic inputs at runtime, parameterized jobs make duties extra versatile and reusable. We handle jobs by means of our CI/CD pipelines, producing YAML information to configure giant exams effectively and repeatably.
- Observe and log experiment outcomes centrally for environment friendly comparability. This offers our workforce clear visibility into what’s working and what wants tuning, with out duplicating effort. The central logging additionally helps deeper analysis of mannequin habits throughout doc varieties and referral situations.
- Leverage historic information to simulate downstream outcomes and refine our fashions earlier than pushing to manufacturing. Catching points early within the testing cycle reduces threat and accelerates deployment. That is significantly vital given the range of referral types and the necessity for compliance inside closely regulated EHR environments like Epic.
This course of was impressed by our success working with Databricks on our deep studying frameworks. We’ve since tailored and expanded it for our eFax work and enormous language mannequin (LLM) experimentation.
Whereas we use Azure AI Doc Intelligence for OCR and OpenAI’s GPT-4.0 fashions for extraction, the actual engineering accelerant has been the flexibility to run managed, repeated exams by means of MLflow pipelines—automating what would in any other case be guide, fragmented growth. With the unifying nature of the Databricks Information Intelligence Platform, we’re in a position to remodel uncooked faxes, experiment with completely different AI methods and validate outputs with pace and confidence in a single place.
All extracted referral information should be built-in into Epic, which requires seamless information formatting, validation and safe supply. Databricks performs a important position in pre-processing and normalizing this info earlier than handoff to our EHR system.
We additionally depend on Databricks for batch ETL, metadata storage and downstream evaluation. Our broader tech stack consists of Azure Kubernetes Service (AKS) for containerized deployment, Azure Search to help retrieval-augmented era (RAG) workflows and Postgres for structured storage. For future phases, we’re actively exploring Mosaic AI for RAG and Mannequin Serving to boost the accuracy, scalability and responsiveness of our AI options. With Mannequin Serving, we shall be in a greater place to successfully deploy and handle fashions in actual time, guaranteeing extra constant workflows throughout all our AI efforts.
From months of backlog to real-time triage
Finally, the beneficiaries of this eFax resolution are our caregivers—clinicians, medical information directors, nurses, and different frontline workers whose time is presently consumed by repetitive doc processing. By eradicating low-value guide bottlenecks, we intention to return that point to affected person care.
In some areas, faxes have sat in queues for as much as two to 3 months with out being reviewed—delays that may severely influence affected person care. With AI-powered automation, we’re shifting towards real-time processing of over 40 million faxes yearly, eliminating bottlenecks and enabling quicker referral consumption. This shift has not solely improved productiveness and lowered operational overhead but additionally accelerated remedy timelines, enhanced affected person outcomes, and freed up scientific workers to deal with higher-value care supply. By modernizing a traditionally guide workflow, we’re unlocking system-wide efficiencies that scale throughout our 1,000+ outpatient clinics, supporting our mission to supply well timed, coordinated care at scale.
Because of MLflow, we’re not simply experimenting. We’re operationalizing AI in a approach that’s aligned with our mission, our workflows, and the real-time wants of our caregivers and sufferers.
