On Built This Week, AI in healthcare shows up in the back office and the lab more than the exam room. The guests we hosted use AI agents to fill home care shifts, to read clinical trial protocols and rank trial sites, and to model DNA and proteins for drug discovery. All of them work under tight privacy rules. Our own healthcare builds went the other way: personal tools that help a patient understand their own medication and genetics.
At Arya and speaking about AI, we only use AI through our cloud providers. So we work with Amazon and Google.
— Arun Kalaiselvan, Episode 24
How companies we talked to use AI in healthcare
Arya Health: scheduling agents for home health agencies
Arya is a workforce platform for in-home health care. When a nurse calls in sick, an office manager can spend two to three hours finding a credentialed replacement the patient is comfortable with. Arya's eligibility system scores caregivers into tiers, and then an AI phone call offers the top tier the shift, sometimes with a bonus. CTO Arun Kalaiselvan said that lifted revenue about 10% on average, and one customer saw 12 pediatric shifts filled over a weekend that would otherwise have gone unfilled.
Principal architect Anand Chandrasekaran shared two lessons at 15:36 and 18:07: newer models turned many multi-shot prompts into single shots, which matters under five-minute SLAs, and typed inputs with Pydantic AI cut hallucinations 40 to 50%.
Ryght AI: matching clinical trials to sites
Dr. Chadi Nabhan, an oncologist who practiced for about 20 years, explained that trials fail for three reasons: the drug doesn't work, it's too toxic, or the trial never enrolls. Ryght goes after the third. It keeps a digital twin of every site that has run a trial, uploads a protocol of up to 200 pages, and has 13 AI agents read the title, schedule of events, biomarkers and endpoints within seconds. It infers site capabilities from past trials, adds US Census demographics for diversity targets, and tiers the sites.
80% of trials run behind schedule, 50% of sites enroll just zero to one patients, and from a sponsor perspective, time is money.
— Dr. Chadi Nabhan, Episode 34
Converge Bio: foundation models for drug discovery
Dov Gertz treats DNA, RNA and proteins as text and trains transformer models on them. Biotech and pharma customers use three systems: antibody engineering, protein yield optimization for manufacturing, and a virtual cell for finding new targets and stratifying clinical trial patients. His pitch is "faster, cheaper, better," and he expects cancer to benefit first because it has the most data.
But can you actually improve your quality of life and length of life? I think those are the real questions that we should be asking ourselves.
— Dov Gertz, Episode 32
Neurable: brain sensing in headphones
Dr. Ramses Alcaide says Neurable's AI boosts the performance of brain data enough to bring brain-computer interfaces out of the lab and into headphones. Besides focus tracking, the app reports brain health metrics such as cognitive strain and brain age. His motivation is medical: he said the standard of care often detects Alzheimer's and Parkinson's about ten years into the disease, and everyday sensors could catch problems sooner.
What we built
- AI start-of-care packet generator: Sam turned an unstructured hospital discharge summary for a heart failure patient into a care plan, a four-week schedule by discipline and matched clinicians. He has no healthcare background; a ChatGPT-written PRD went into Google AI Studio and took about 30 minutes. Arya's team said it got the disciplines right and already showed something planned for their fourth version.
- Clinical trial site selection dashboard: Sam's simulation of manual versus AI-optimized site selection, with biomarker and census toggles. Dr. Nabhan said Sam "certainly picked up on what we're doing."
- Cholesterol medication 3D visualizer: Jordan, who has genetically high cholesterol, built a rotating 3D liver and intestine model in Claude where you toggle Crestor, Zetia and Repatha and watch LDL drop. He pictures doctors using one to explain a prescription.
- Personal DNA and health insights dashboard: Jordan gave Claude Code his raw 23andMe file, blood work and medications, kept everything local in SQLite, and had it write a report for his sister to take to her doctor. Not every result held up: it said he was lactose intolerant, and he isn't.
What's still hard
- Privacy narrows your model choices. Anand said few AI tools are HIPAA compliant enough for patient data, so Arya only uses models its cloud providers offer and has had to push Bedrock to add a specific version.
- No memory means repeating everything. Arya sends no long-term memory with prompts, so every call carries all the exclusions and context.
- Costs can jump. Arun said one AI bill came in about six times higher than expected.
- Prototype to production is long. Sam built his care packet in 30 minutes, but said there are "a million steps in between" concept and real use. Anand's one gap: in practice the process isn't real time.
- Trials stay slow. Dov expects AI to shorten discovery first; he assumes the five-year clinical trial bottleneck will remain for the foreseeable future.
- Thin labeled data. Dov puts molecular AI five to ten years behind text models.
On patients asking chatbots about symptoms, Dr. Nabhan was clear at 24:26: an informed patient and family is "10 times better" than an uninformed one.
Episodes on AI in healthcare
- Episode 13: Jordan's 3D cholesterol medication model (September 2025).
- Episode 24: This AI Fills Healthcare Shifts in Minutes with Arun Kalaiselvan and Anand Chandrasekaran of Arya Health (December 2025).
- Episode 30: Jordan's local DNA and blood work dashboard (February 2026).
- Episode 32: From DNA to Drugs with Dov Gertz of Converge Bio (February 2026).
- Episode 34: AI Is Rebuilding Clinical Trials with Dr. Chadi Nabhan of Ryght AI (March 2026).
- Episode 41: Dr. Ramses Alcaide demos Neurable's brain-sensing headphones (May 2026).
FAQ
How is AI used in healthcare today?
The companies we hosted use it on operations and research: Arya Health's AI agents schedule home care staff and phone caregivers to fill open shifts, Ryght AI matches clinical trials to the sites most likely to enroll patients, and Converge Bio trains foundation models on DNA, RNA and proteins for drug discovery.
Can you use ChatGPT or Claude with patient data under HIPAA?
Arya Health's CTO said they only call models through their cloud providers, Anthropic via AWS Bedrock and Gemini via Google Vertex AI, because those contracts cover data at rest and in transit. They also send no long-term memory with any prompt.
Will AI make drug development faster?
Converge Bio founder Dov Gertz expects the biggest time savings in discovery and preclinical work. He assumes the roughly five-year clinical trial phase won't change much for the foreseeable future, though he predicts better success rates.
Is it good or bad that patients use ChatGPT for medical questions?
Oncologist Dr. Chadi Nabhan of Ryght AI called it a net positive: an informed patient and family is better than an uninformed one, and the doctor's job becomes separating signal from noise.