During one of my consulting contracts, I’d round at my physicians office once a month or so and I observed that their billing staff were consistently struggling to get all the information needed into one single place in order to bill accurately and in a timely manner.
Furthermore I observed that the method itself of anesthesia professional billing consisted of humans executing a system of well-documented rules to translate clinical documentation into CPT and ASA codes. My best friend at the time was just teaching himself some early ML techniques and we decided together that we could build a rules engine if we had a large enough dataset of clinical documentation and historically billed encounters.
I worked with a SQL engineer to write and run an extract daily, and with my cofounder to iterate on the model. Working together, and with an early form of RLHF, we compiled a rigorous algorithm capable of processing 50% of anesthesia encounters at >99% accuracy. We signed up that first client, and I served as Product Lead compiling it into a suite of enhancements that offered value working with some offshore development resources.
We scored an interview with YCombinator, but they were concerned (rightly) that enterprise healthcare sales would be difficult to secure tangible progress on the time horizon they preferred to focus on. We bootstrapped for a few years, but healthcare sales is hard and we were unable to secure additional clients - but we’re immensely proud of the product we built.