Before Episode 24 (December 12, 2025), Sam built a prototype for our guests at Arya Health: paste in a hospital discharge summary, get back a home health start-of-care packet. He has no healthcare background. The first working version took four or five prompts over about 30 minutes in Google AI Studio, followed by small refinements.
On our prep call with Arya Health, start-of-care intake came up as a feature worth building. When a patient is discharged from the hospital, a home health agency has to turn an unstructured, somewhat coded summary into a plan and a staffing schedule. Sam wanted to see how far AI could get without domain knowledge, and said up front at 02:22 that it "could be a little bit off."
Got a PRD from ChatGPT, went to the Google AI studio using Gemini Pro, put the PRD, and then within four or five different tweaks, probably in the span of thirty minutes, I had kind of what we're seeing now.
— Sam Nadler, Episode 24
The demo input was a complex congestive heart failure (CHF) discharge summary with shorthand like "2W3, 1W4" that Sam admitted he couldn't read. The packet it generated had a diagnosis and risk factors, a clinical strategy, a projected four-week schedule by discipline, matched staff (a skilled nurse with wound care skills who speaks Spanish), an auto-drafted outreach message, home context (four steps to enter, an elderly wife who may not be able to lift the patient), missing info to follow up on, and clickable workflow actions.
Arya Health CTO Arun Kalaiselvan called it "a pretty good starting point" and said it got the titles right: skilled nursing, occupational therapy, physical therapy and home health aide. Principal architect Anand Chandrasekaran went further:
But what's most interesting in this is something that we are hoping to do in, the fourth version is already captured in this.
— Anand Chandrasekaran, Episode 24
His one correction: the real process isn't real time, and the demo implies it is. The outreach button didn't actually send anything; it's a prototype.
Sam's own takeaway was that AI gets you "from, like, zero to one so fast," but there are "a million steps in between" a concept and real use. He didn't know how discharge notes actually arrive. Jordan added the bigger blocker at 08:58: patient data can't go to a typical Supabase or Firebase project like other vibe-coded apps. It has to be HIPAA compliant, which is what the rest of the episode covers. More on that in AI in healthcare.
As a prototype, yes. Sam's tool reads an unstructured discharge summary for a complex CHF patient and produces a diagnosis and risk summary, clinical strategy, a four-week schedule by discipline, matched clinicians and missing-info flags.
Sam researched the company, had ChatGPT write a PRD, pasted it into Google AI Studio with Gemini Pro, and made four or five tweaks over about 30 minutes, followed by small refinements to add features.
No. Jordan pointed out you can't push patient data to a typical vibe-coding back end; it needs HIPAA compliance. Arya's team also noted the real start-of-care process isn't real time the way the demo is.
In Sam's version: diagnosis and risk factors, clinical strategy, a projected four-week schedule for skilled nursing, OT, PT and home health aide, matched staff with competencies and languages, home context, missing info and workflow actions.
