To our knowledge, the only approved app on the App Store that allows users to run an LLM on medical records gathered via HealthKit (this only works for large providers in the US, UK and Canada).
Yari Timeline allows you to build an organized, accessible, and private timeline of your care. It runs a HIPAA-compliant LLM (Gemini on Vertex AI) on the HealthKit medical records in three careful steps: (1) it generates a global timeline of all records, one episode of care at a time; (2) it generates a summary of each episode in plain language for laypeople to understand; (3) it allows the user to go all the way to the individual medical records on their HealthKit and see their AI summary or the raw data.
We purposefully built this app NOT as a chatbot. This allows us to thoroughly validate its prompts and context-assembly on a predefined set of use cases and build a user experience that minimizes common failures.
Having said all that, the app is not intended for medical use and can be wrong. It’s merely a demonstration of what’s possible, done in perhaps the most careful and private way.
As a bonus, we never store any data locally or on our servers. We have even disabled Gemini’s caching capabilities and traffic logs. You are always in incognito mode :)
It’s worth saying that the experience is far from ideal. Given the strict constraints on Apple HealthKit’s medical records, we cannot do things like storing pre-processed data in the app to reduce initialization time. We also cannot provide ways to do data sharing or long-running processing. You also may end up in a bit of a messy state if you do not give the app access to all records during the HealthKit authorization workflow (you can fix this in iOS Settings). So please make sure to follow the instructions carefully and feel free to write to us via [email protected] if you run into any issues.
This app demonstrates what’s possible with today’s App Store rules and today’s technology. But it’s just scratching the surface of what’s possible. For example, we are eager to add a lot of interesting new things once iOS 26 comes out and on-device models are accessible to more users.
A short, messy, and honest story of my personal life and career (as a former member of the initial Bard team). My time at Google coincided with a brush with cancer, which ultimately led me to quit and pursue something more meaningful. I’m now working on a startup focused on improving care quality for cancer patients. Writing this made me feel incredibly vulnerable, but I hope it opens the door for honest conversations about the existential challenges of working in the tech industry.
In the spirit of the authors' point of sharing one's supporting evidence, could you please share the data and code for scraping the list of devices from the FDA website? :)
TL;DR: I explored the journey of fine-tuning two small LLMs: Llama 3.1-8B-Instruct vs GPT-3.5-Turbo on a fun dataset of GPT-4o generated conversations in the style of Larry David. All the code for this project and evaluation results are openly accessible.
My key takeaway was that fine-tuning small LLMs is easily within reach and practically free for independent developers. This approach could provide a major cost/latency/privacy advantage if one is interested in small models for specialized use cases.
I stitched up Twilio, LLMs, speech understanding, and speech generation tools together to work all in parallel in a system. The result is a low-latency, interruptible and near-human quality AI phone agent
Wanted to share my learnings and code with the community. Hope this ends up being useful for someone out there :)
Location: SF Bay Area
Remote: No
Willing to relocate: only to nyc
Résumé/CV: https://www.linkedin.com/in/akiani/ - can send resume on email. just don't want it to be scraped here :)
Email: kiani [ at ] cs.stanford.edu
I am a very technical product person with years of hands on work in software/ML. My passion is in healthtech. Ex-Googler, acquired startup 1st employee, Stanford CS masters.
Yari Timeline allows you to build an organized, accessible, and private timeline of your care. It runs a HIPAA-compliant LLM (Gemini on Vertex AI) on the HealthKit medical records in three careful steps: (1) it generates a global timeline of all records, one episode of care at a time; (2) it generates a summary of each episode in plain language for laypeople to understand; (3) it allows the user to go all the way to the individual medical records on their HealthKit and see their AI summary or the raw data.
We purposefully built this app NOT as a chatbot. This allows us to thoroughly validate its prompts and context-assembly on a predefined set of use cases and build a user experience that minimizes common failures.
Having said all that, the app is not intended for medical use and can be wrong. It’s merely a demonstration of what’s possible, done in perhaps the most careful and private way.
As a bonus, we never store any data locally or on our servers. We have even disabled Gemini’s caching capabilities and traffic logs. You are always in incognito mode :)
It’s worth saying that the experience is far from ideal. Given the strict constraints on Apple HealthKit’s medical records, we cannot do things like storing pre-processed data in the app to reduce initialization time. We also cannot provide ways to do data sharing or long-running processing. You also may end up in a bit of a messy state if you do not give the app access to all records during the HealthKit authorization workflow (you can fix this in iOS Settings). So please make sure to follow the instructions carefully and feel free to write to us via [email protected] if you run into any issues.
This app demonstrates what’s possible with today’s App Store rules and today’s technology. But it’s just scratching the surface of what’s possible. For example, we are eager to add a lot of interesting new things once iOS 26 comes out and on-device models are accessible to more users.