Arya Health's CTO and principal architect on AI agents that call caregivers to fill open shifts, HIPAA-safe model use, and why Pydantic AI cut hallucinations.
Arya Health fills open home healthcare shifts by letting an AI agent find the right available, credentialed caregiver and call them directly, a job that can take a human admin two to three hours. That is how Arun Kalaiselvan, co-founder and CTO, and Anand Chandrasekaran, principal architect, described Arya on Episode 24 of Built This Week. Arya is a workforce management platform for in-home healthcare that also handles onboarding, credentials and compliance, so office managers can focus on revenue generating work.
Before the interview, Sam built a prototype with Arya in mind: a start of care packet generator. It takes a hospital discharge summary for a congestive heart failure patient and produces risk factors, a clinical strategy, a projected four week schedule by discipline, and a list of matched staff, such as a skilled nurse with wound care skills who speaks Spanish. It also flags missing information and drafts outreach to the matched clinicians.
Anand said it captured the workflows well, including something Arya hopes to have in its fourth version. The one thing he thought was off: in real life a lot of that intake is not real time. Arun said it got the disciplines right, from skilled nursing to occupational therapy, physical therapy and home health aides.
Sam's workflow was simple. He researched Arya, got a PRD from ChatGPT, and built the demo in Google AI Studio with Gemini Pro in about thirty minutes and four or five tweaks. His own caveat: going from a concept like this to real world use takes a million steps in between.
Watch Arya Health on HIPAA-compliant AI in Episode 24 (starts at 9:52) · Episode 24 notes and transcript
Arun called this the most complicated part of the whole ecosystem. He says about 70% of Arya's infrastructure is hosted by Arya itself. The parts that are not, mainly phone calls and SMS through vendors such as Bland, Twilio and Amazon Pinpoint, are covered by agreements specific to HIPAA. Each customer's data is stored separately, and he says data never leaks from one customer to another.
For AI, Arya only uses models through its cloud providers. It talks to Anthropic through Amazon Bedrock and to Gemini through Google Vertex AI, and those contractual relationships cover data at rest and in transit. Anand added that Arya is multi cloud, with a separate peering network between GCP and AWS so the clouds only talk through that tunnel. He checks new models constantly, but Arya's workflows only use what its cloud providers offer, and the team has pushed Bedrock to make a particular model version available.
We cover more companies working under these rules on our AI in healthcare page.
Jordan asked what the latest models made possible. Arun started with cost, because a change on Anthropic's side had just surprised them:
Our last bill was about six x for what we would expect it to be because there was a change in Anthropic and how much time it took our Lambda to actually have that connection open with Anthropic.
— Arun Kalaiselvan, Episode 24
The calls behind it were simple ones that write a human readable description of a shift to text to caregivers. Anand focused on speed. Arya sometimes has five minute SLAs with customers, and in his analytics many tasks that used to need multi shot prompts are now single shot. Prompts can be more precise without piles of examples, and he called Gemini 3 and Gemini 2.5 Flash very fast.
Arun pushed back on one point. Arya allows no long term memory in any prompt it sends, so every query still has to carry all of its exclusions and context.
Anand's tip for builders was Pydantic AI. Its dependencies work like an input type, and he says that for 90 to 95% of everyday agent tasks he no longer needs to give examples. He uses typed inputs and flags to steer the agent's thinking with a structure it understands, and moved everything into Pydantic classes instead of JSON types.
After starting to use that, it's been extremely fast development and the hallucinations have reduced 40 to 50% for me just by changing the input types with Pydantic AI.
— Anand Chandrasekaran, Episode 24
He recommends it especially for single agents. Jordan said he would download the library right after the show.
Scheduling is Arya's foot in the door. Arun's example: you manage 100 nurses and one calls in sick. Someone has to find that nurse's shifts, then find people who are available, hold current credentials and work well with the patient, plus a lot of tribal knowledge. That can take two to three hours, and you still don't know if the person will say yes.
Arya already knows who is available, where they are, their credential status and whether they can work with that patient. AI calls them, tells them a patient's shift just opened, offers a bonus and gets the shift filled. Arun says this has added close to 10% in revenue on average for existing customers. His favorite example came from a customer's CEO: over one weekend, Arya filled about 12 pediatric shifts that he says would otherwise have gone unfilled.
Sam pointed out that more consistent care should mean better care. Arun agreed, and said the team changed how it talks about the work:
So we've stopped talking about shifts. We're talking about people now.
— Arun Kalaiselvan, Episode 24
Instead of posting an open shift, Arya says a patient needs care, for example every Monday, and a caregiver can take that patient for the rest of their care period. The goal is care teams, not filled slots. Anand said what makes it work is a heuristic eligibility system, one of the first things they built, which scores caregivers into tiers per patient. Urgent calls go to the top tier first. Arun says they are now looking at drive time as a traveling salesman problem, which Jordan noted becomes one with time windows.
In the news segment, Jordan said Cursor is his go-to IDE, and that he is surprised how good Claude Code is. He argued that model makers may end up with the best tooling. Anand expects coding models to saturate within two years, at which point he thinks tooling and price decide the winner, and Cursor could be well placed. He also thinks there will always be two or three winners plus open source.
Arun was more skeptical. He says Arya ran a six month enterprise pilot with Cursor and found it introduced more issues than it solved, so they switched to Copilot; Anand noted that was over a year ago and Cursor has matured since. Arun's bet is on Google:
I do think Google is going to be the silent killer in this current ecosystem.
— Arun Kalaiselvan, Episode 24
He credits the Vertex AI team's work over the last three years and says Arya is all in on Gemini. Learn more about Arya Health.
Arun and Anand also talk about AI dating apps and their own hackathon matching engine from their NYU days. Watch Episode 24 of Built This Week, where the full transcript is also on the page.
