A few months ago, we passed the one-year mark of building Counsel.

Over the past year, we’ve grown from a scrappy three-person team to a full-fledged medical group serving patients across the country. I wanted to jot down a few takeaways from building AI systems in healthcare, along with some reflections on the journey so far.

Counsel’s Mission

Counsel's mission is to multiply the world's healthcare capacity. That means a few things:

• We're building a new kind of medical group from the ground up, specializing in asynchronous care delivery.

• We're embedding AI tooling throughout our custom-built clinician cockpit to maximize efficiency and quality at every part of the clinical stack.

In the process of building, I’ve picked up a few technical takeaways:

Vanilla RAG is not enough for medical data

A core component of Counsel's platform is effective retrieval of a patient's medical records, so our physicians can provide extremely personalized medical advice. In our initial iterations of RAG, we relied primarily on semantic similarity-based techniques for search and retrieval. We quickly discovered that not only is medical data incredibly messy, but oftentimes, semantic similarity is not a good metric for clinical relevance.

Measuring clinical relevance requires factoring in aspects like time decay, ontological relationships, and a host of other nuances that clinicians are trained to recognize. We've found that incorporating a mixture of retrieval techniques that account for these factors yields much better results.

The interface is crucial

Something that surprised me about building effective AI systems is that designing the right interface is just as important as improving model outputs. Poorly thought-out interfaces can limit the real-world effectiveness of otherwise powerful models.

This is particularly true when integrating AI into complex human workflows, such as clinical care, where multiple intermediate touchpoints between humans and AI are necessary.

At Counsel, because we have full control over our care platform, we've been able to design novel UIs that optimize how clinicians use AI-generated outputs. A recent interface redesign of our clinician cockpit resulted in a double-digit improvement in time-to-resolution, even while holding model output quality constant.

Measure the right things

On a related note, measuring raw model performance alone is oftentimes not enough when the end goal is increasing human efficiency. It doesn't matter how high your model's F1 score is if the outputs don't drive meaningful improvements in workflow efficiency or care quality.

We've found that tracking metrics such as click-through rates on AI suggestions, edit distances, and end-to-end clinical simulations have all been effective ways to measure the model's true impact. The key is incorporating these metrics into a robust training and evaluation pipeline, which is part of the ongoing work I'm most excited about.

Bringing "software copilots" to clinical care

While copilots now exist in every industry, the level of integration that software copilots have achieved is on another level. If you've used coding tools like Cursor, you know the magic of having a single keystroke take you to the exact line of code you were about to edit, or automatically retrieve the specific files you needed for writing a function. These copilots are embedded seamlessly within the underlying workflows, which is partly why software development has seen the greatest efficiency gains from using AI tools.

At Counsel, we're building copilots that bring that same level of integration into clinical care. Some interesting questions we've been thinking through:

• What action primitives form the building blocks of a clinical encounter?

• What does tool-calling for a physician look like?

• How do we iteratively retrieve context and manage agentic state throughout the course of a complex medical issue?

When we fully control the interface and the workflow of clinical care, these are some of the exciting questions that suddenly become unlocked.

On a non-technical note:

Medical care is a superpower

One of my favorite aspects of building at Counsel is that our engineering team sits in the same room as our physicians, who are providing medical care to patients every day. As a result, I'm constantly in awe of my physician coworkers and what they do. Medical care is truly a superpower. As engineers, it’s rare for our work to have the kind of direct and fundamental impact on a person's well-being as a physician's work does. However, building Counsel does sometimes feel a bit like tapping into that superpower, or at least playing a part in distributing it more efficiently into the world.

If you've read this far, chances are you may be interested in the kinds of problems we're trying to solve. If that's the case, you should consider joining us!