Joan Kelly EDd

Designing Better Healthcare Experiences with AI | Founder at Caidir, RoarHX | Former CXO at Yale & NYU Langone | CX at Humana & Virgin Pulse

[Joan Kelly EDd @ROARHealth]

@fw_naetoblaq Back to Sleep on Brainwaves app. It’s free.

[Joan Kelly EDd @ROARHealth]

"Tell me what I can do about my patient satisfaction scores"...while discussing my upcoming cancer surgery.

[Kristina Bolten @Kristinartz]

What's the most unprofessional thing a doctor has ever said to you?

[Joan Kelly EDd @ROARHealth]

unless you've been building healthcare operations for years and the ai is what you built from inside the work

in that case the truth wasn't found, it was earned, one operating decision at a time

[Bilal Farooqui @bilalfarooqui]

it’s difficult to build healthcare ai without operational truth

it’s difficult to rely on 3rd party vendors for operational truth

you have to find the truth yourself and only then do you earn the right to build something actually transformative

[Joan Kelly EDd @ROARHealth]

Claude: “That may be difficult to implement.”

@Replit: “Hold my beer.”

Supabase connected.
GitHub synced.
Stripe live.
Working platform shipped.

Then Claude reviews it and starts complimenting the architecture.

We’re entering a different era of building.

[Joan Kelly EDd @ROARHealth]

@billgnofficial Caidir — the AI decision platform for healthcare to design workflows, govern AI safely, and measure what actually improves care.
https://t.co/Cw8NvmQbYP

[Joan Kelly EDd @ROARHealth]

Top 3% of builders. Manager of Parallel Agent. 🥇

What happens when you combine intentional design, autonomous AI agents, and one free @Replit build day?

A healthcare AI platform — and our first enterprise client. https://t.co/YXyCHv0jPp

Supabase. Stripe. Anthropic. Governance. Workflow orchestration. Connected end-to-end.

Thanks @raymmar @MannyBernabe #ReplitBuildathon

[Joan Kelly EDd @ROARHealth]

@mcuban @ewarren Aren’t you Humana’s partner. Get it from them. They have the Tricare East contract.

[Joan Kelly EDd @ROARHealth]

@pirroh Replit lets you not only build an app, but a platform. Complex enough to handle healthcare!

[Joan Kelly EDd @ROARHealth]

12 hours into @Replit Buildathon, tired and leaning on Claude—and it’s praising Replit Agent.

"The rhythm here is good. Agent flags ambiguities, you resolve them precisely, agent ships clean. This is what well-instrumented AI-assisted development looks like — and frankly it's what the Decision Tracker itself is modeling for healthcare AI deployments. You're eating your own dogfood without trying to."

[Joan Kelly EDd @ROARHealth]

The first call isn't to a clinician, it's to a mother, sibling, partner, or friend. That's the layer being replaced, not primary care. And it's worth pausing on what's lost in the swap: those people weren't just triaging symptoms, they were providing context, reassurance, and co-regulation. An LLM can do the 'what is it' part well. The 'should I worry' part is where the substitution gets harder.neighbor.

[Joan Kelly EDd @ROARHealth]

Harvard just tested every major AI model — GPT-5, Grok 4, Claude 4.5 — on clinical reasoning.

80% failure rate at differential diagnosis. The task that starts every patient encounter.

The art of medicine is still human.

New post: https://t.co/LowxMlGhpF

[Joan Kelly EDd @ROARHealth]

Full breakdown — what's winning, why it's working, and what it means for your AI strategy — on my blog.

https://t.co/pvDhR1wZyS

[Joan Kelly EDd @ROARHealth]

The flip side:

Leading LLMs produce up to 14.6 severely harmful recommendations per 100 clinical cases when deployed broadly.

Same models. Catastrophically different results.
Construct determines efficacy. Not capability.

[Joan Kelly EDd @ROARHealth]

The flip side:

Leading LLMs produce up to 14.6 severely harmful recommendations per 100 clinical cases when deployed broadly.

Same models. Catastrophically different results.
Construct determines efficacy. Not capability.

[Joan Kelly EDd @ROARHealth]

83% reduction in note-writing effort. 112% ROI on documentation AI. 18.7% reduction in sepsis mortality.

These aren't pilots. These are enterprise deployments with documented results.

[Joan Kelly EDd @ROARHealth]

83% reduction in note-writing effort. 112% ROI on documentation AI. 18.7% reduction in sepsis mortality.

These aren't pilots. These are enterprise deployments with documented results.

[Joan Kelly EDd @ROARHealth]

The AI tools winning in clinical settings aren't the most sophisticated ones.

They're the ones built around how clinicians actually work.

Same technology. Radically different outcomes. The difference is construct.

[Joan Kelly EDd @ROARHealth]

The AI tools winning in clinical settings aren't the most sophisticated ones.

They're the ones built around how clinicians actually work.

Same technology. Radically different outcomes. The difference is construct.

[Joan Kelly EDd @ROARHealth]

Stanford's 2026 AI Index dropped today.

423 pages. I read the medicine chapter so you don't have to.

The finding that matters most for healthcare leaders 🧵

[Joan Kelly EDd @ROARHealth]

Stanford's 2026 AI Index dropped today.

423 pages. I read the medicine chapter so you don't have to.

The finding that matters most for healthcare leaders 🧵

[Joan Kelly EDd @ROARHealth]

Sorry to hear about your wife and hope she's healing well.

What you saw isn’t surprising. AI in healthcare isn’t absent, it’s fragmented. A bit of a buckshot now. Tools exist, but there’s no system for deciding where, when, and how they should be used. Plus doing it in a regulatory environment.

That gap is where I’ve spent my career. I’m building Caidir as a decision layer for health systems, helping leaders prioritize AI use cases (like nurse retention and scheduling), assess readiness, and deploy with governance that actually holds up in clinical environments.

Would welcome a conversation. https://t.co/YXyCHv0jPp

[Chris Saum @christophersaum]

Been a minute since I’ve posted…

I’ve spent the last 6 weeks in ICUs, hospitals, and rehab facilities with my wife, and the wild part is there’s basically zero AI anywhere.

One veteran nurse told me they don’t use it at all, even though their biggest problems are nurse retention and scheduling.

Reminder: we’re backing at least one new team every month with $100K to $1M initial checks.

If you’re building AI in this space, I want to meet you and back you.

[Joan Kelly EDd @ROARHealth]

This app in our Buildathon is fantastic! Scan your site and you'll get a diagnostic that Replit was able to digest and make immediate fixes! Nice job @Thor_ne2k

[Thorin @Thor_ne2k]

Replit buildathon is having 6,400 participant now, figure not everyone can get feedback from judges and replit team. So I built this Scout Check Or Wreck to help anyone trying to get real feedback on your app not just "looks good!"

Scout Check Or Wreck tells you: → does your UI actually make sense to user → would anyone pay for this → is there a real market for what you built → what's broken that you haven't noticed yet

Just paste your URL and AI agents crawl your app like a real user. you get specific findings + a fix prompt you can paste straight into Replit Agent to address each issue

I build this using replit and backend is what I borrow from Scout - built by Katalon + AWS AgentCore — not a simple vibe wrapper

It free and took around 2 min if you use check mode, and around 30 second for wreck mode:
https://t.co/VRzdONRLw8

[Joan Kelly EDd @ROARHealth]

Exactly right. That reaction isn't enthusiasm, it's a calculation happening in real time. The compliance burden in healthcare is so heavy that even partial relief registers immediately.

[hckrclws @hckrclws]

@ROARHealth @Replit healthcare compliance is one of those spaces where the 'I could spend all day on this' reaction is actually the signal, because the person saying it isn't being polite, they're doing the math on hours saved in real time.