community care team

Inspiration

By 2030 there will be an 85K physician shortage in the US to care for the 80+MM seniors navigating chronic diseases like diabetes, hypertension and COPD. Their quality of life is at risk with the quality of care, and it is estimated that this gap in care will cost the US $500B in extra Medicare spend.

Heartbreaking and terrifying.

communitycareteam + Perplexity

We're going to fill the 85,000 doctor hole.

The infinite potential of Perplexity is its ability to power tools that upskill someone with its deep research and meticulous citations. For healthcare, it means that in specific situations a caregiver can deliver at two to three levels above their traditional training.

Picture Margaret, one of the million plus retired nurses, sitting at a table in one of the 20,000 community centers across the US. Bob, a senior she has been working with for a couple of months sits down, maybe gets a quick glucose check or has his blood pressure taken. And then our "upskilled" Margaret takes him through a ten minute review of his current health numbers and care plan.

We're going to fill the 85,000 doctor hole

The next day Bob logs in to community*care*team, gives his care plan a quick look, records his latest weight, and then is served up three articles to help him build his habits in diet, physical activity, taking his meds and writing important stuff down. Bob is an increasingly more pro-active participant in his health.

We're going to fill the 85,000 doctor hole

the four coolest ways we used Perplexity

careplan	conversation	questions	education

With input from the American Diabetes Association, American Heart Association and the Mayo clinic, combined with longitudinal synthetic patient data, Perplexity creates individualized care plans with priorities, goals, and medically reviewed inspiration	It's sometimes hard to talk about health. It's a little scary. It's a little personal. Perplexity assumes the "voice" of communitycareteam and creates a script for the caregiver to lead a discussion at the right level of compassion and simplicity to make an impact.	The best discussions are two way, and another way communitycareteam "upskills" the caregiver is by putting a simple open-question box inside the script. Of course we have to put guardrails in place to keep everyone safe	We have 19 categories of knowledge from the ADA, AHA, Mayo and Cleveland Clinic. We have 300 prompts across those categories for patient education. Perplexity takes the patient priorities, the categories and prompts and builds daily education for the patients.

Making the Cake

We used a variety of ingredients. Simple javascript on the front end and AWS on the back end to do the heavy lifting. Cognito to manage identity, RDS for data storage, API gateway to bridge the front- and the back- and Python+Lambda to make the magic.

It wasn't all as easy as we made it look...

We spent a lot of time iterating on extracting the right "voice" from Perplexity. Friendly, but not goofy. Personal, but not too intimate. Structured, but not too stiff. We got about 80% to where we wanted. We also couldn't figure out in education development how to get Perplexity to be influenced by something, but not just straight copy it.

We also had to figure out how to navigate real time calls to Perplexity in an environment of 30 second API timeouts and a desire for rapid response.

Aside from that it was mostly the frustrating friction common to all dev projects like "why isn't the matplotlib making the chart like I expected" or "why isn't the cookie ID persisting?" or "why did that last Perplexity response sound like something from Gone with the Wind?"

A lot of it worked out great

The care plans ended up being a lot like would have designed ourselves just based on experience. We also figured out a couple of ways to make requests of Perplexity before we needed it so that when we were ready, the work was already done,

What we're most proud of is that we pushed past the novelty of this super powerful brain machine that is Perplexity where you can imagine a million things you could do, but most of them you only want to do once or twice. We listened to a couple of doctors and patients to understand "what would you accept in mid 2025 from AI" and really pursued the promise of AI as an enhancement, not a replacement, a tool for "upskilling", not "displacing".

What we learned

We did a handful of physician and patient interviews before digging in our mission. The output was in the gray zone between "informed" and "anecdotal", but there were some themes

physicians are looking to AI to a) solve for corner cases, b) give them more time, c) find them more money
docs are hesitant to rely on some AI conclusions because they are still liable if anything goes wrong
someone else picking up between-appointment-care gives them more time. Also, they are willing to share non-critical activities like improving adherence or building habits.
they hate AI that tells them what everyone already knows. When it says that a lifetime smoker has an elevated chance at COPD, they roll their eyes
patients don't lack information, they lack structure and organization. Outside of crisis, people are not information seekers, health info has to come find them. There is less a sense of "I don't know enough" and more "I'm not sure what I need to know, or how to organize it, or how to prioritize it". My favorite quote was "my doctor's great at telling me what to do. She's terrible at telling me how to do it."

What's next for community care team

There's a long road between a hackathon product and something we would want people to use to manage their health data and health priorities. There's also 100 UX moments we're not 100% happy with. What's next is really finding out who might be the caregiver and who might be the patient, beyond our market speculation. Also, we got enough variance in some of our responses from Perplexity that we need to figure out some LLM-specific unit testing to make us feel like we have a better handle on expected output from the bit brain machine.