I’m Joshua, a student, and I’m excited (and a little nervous) to share something deeply personal that I’ve been working on: Islet, my diabetes management app powered by GPT-4o-mini. It’s now on the App Store, but I want to be upfront—it’s still very much in its early stages, with a lot more to go.
I was diagnosed with Type 1 diabetes while rowing competitively, and that moment changed everything. It wasn’t just the practical challenges of managing insulin, carb counts, and blood sugars; it fundamentally shifted how I see myself and the world. It forced me to slow down, prioritise my health, and take control in ways I never had to before. My outlook on life became more focused on resilience, adaptability, and finding solutions to problems that truly matter.
This app started as a pet project over the summer, a way to see what I could create using ChatGPT and explore the potential of LLMs to help with real-world challenges. At first, it was just about making my own diabetes management easier—understanding patterns in blood sugars, planning meals, and adjusting routines. But as I worked on it, I realised it could do more.
Right now, Islet offers personalised meal suggestions, tracks activity, and provides basic insights based on the data you enter. It’s far from complete. Even so, the process of building Islet has already taught me so much about how powerful AI can be in creating personal, meaningful tools.
This project is deeply tied to how my diagnosis changed me. It’s about more than managing diabetes, it’s about showing how anyone, even a student experimenting over the summer, can use AI to potentially solve real, personal problems. I believe tools like LLMs have the power to democratise solutions for all, making life just a bit easier for all of us.
If you’re curious, you can check it out here: https://apps.apple.com/gb/app/islet-diabetes/id6453168642. I’d love to hear your thoughts what works, what doesn’t, and what features you think would make it better. Your input could help shape the next steps for Islet.
Thanks for reading !
joshua
I downloaded the app, just to check it out, and the one thing that just struck me right off the bat is the permissions. Read is fine, it’s the write permissions, particularly glucose level and insulin delivery. I don’t know the full app ecosystem, and if it would be possible for your app to interfere with the insulin delivery settings on her pump.
if you have a website or anything that discusses how the application operates and what the permissions are used for it definitely check it out.
Ultimately though this looks like a great tool!
(Also really digging the company name!)
It looks like the intentions are good, but it reminds me of some indie hacker with zero AI (apart from wrapper apps) or therapist background offered “therapist AI”.
Some people don’t understand how much garbage AI outputs, and these people might not be skeptical enough when it comes to taking medical advice from gpt
I would be really careful in this area though, especially using ChatGPT to generate suggestions. This to me this does venture into medical device territory, based on the intended use. Check the guidelines here https://www.gov.uk/government/publications/medical-devices-s... - UK specific, but will be similar for FDA.
Honestly, I would seek proper consultancy advice, remove any suggestions / recommendations for now, and just have it as a data-logging platform. The disclaimer unfortunately will not stand up.
Congratulations on getting this far - I really hope you continue on this path, just make sure you are on firm ground.
Especially when the problematic features are charged for, it gives the recommendations / suggestions an air of legitimacy which could be dangerous.
So you either get lucky and your doctor can prescribe you a commercial loop, or you compile one from source.
I work (freelance) with a consultancy [1] that helps specifically with software-as-a-medical-device (SaMD). My email is in my profile if you want to chat about what might be the best way forward.
I am not sure how ChatGPT can give any advice (if it is even given in the app) about insulin or glucose.
The Islet app is designed as a knowledge base that logs crucial data, including insulin dosages, meals, and physical activities. It aims to provide users with insights into how these factors interact to impact their blood sugar levels. Here's a breakdown of how ChatGPT integrates into the app and what data is involved:
What kind of data powers ChatGPT in Islet?
The ChatGPT component in Islet acts as a translation and query layer rather than the sole knowledge source. Islet’s knowledge base aggregates and organises the user’s logged data, such as:
- Glucose Levels: Derived from CGM (Continuous Glucose Monitor) data. This Data is currently on a 3hr delay in the app.
- Insulin Dosages: Logged by the user to capture the timing, type, and amount of insulin administered.
- Meals: Users can log meals in detail, including macronutrient composition, portion sizes, and timing.
- Activities: Logs include exercise type, intensity, and duration, as physical activity significantly impacts glucose regulation.
Does ChatGPT provide advice?
While the app itself does not directly provide medical advice, the ChatGPT integration facilitates better use of the logged data by enabling the user to ask targeted questions. For example:
- "How has my blood sugar been affected by pasta meals over the past two weeks?"
- "What impact does my 30-minute cycling routine typically have on my glucose levels?"
- "Are there patterns between my evening meals and morning glucose levels?"
Purpose of the App
The primary purpose of Islet is to empower users with a system that captures and organises their diabetes-related data into a knowledge base. The ChatGPT layer makes querying this knowledge base intuitive by translating user questions into actionable insights, offering:
1. Pattern Analysis: It helps users understand trends by analysing recurring meals and activities regimens and their effects on blood sugar levels.
2. Education: Users gain a better understanding of their unique responses to different scenarios, supporting informed decisions in their diabetes management.
By focusing on personalised, data-driven insights rather than generic advice, Islet ensures that the ChatGPT integration remains a helpful tool for exploring user-specific trends.
As a T1D dad it's great to see work like this, not least because it should raise the expectations of what we should expect from well funded engineering teams within the established players.
Right now it's not a great fit for my use case, in that one phone follows my 9 year old around, but the bulk of management happens on other phones on distinct accounts.
I agree with @jcims point about the permissions, greedily allowing write permissions to those data items up front feels like a real point of friction, I'd much prefer to see that done lazily at the point the user is enabling a feature that needs it.
Congrats on shipping, I’m hopefully a month or so away myself :)
Something to consider and probably add some kind of warning about.
I was daily driving it for a few weeks just to see if I wanted to do anything with it, but a flair up and life got in the way.
I have no real plan to release it, but I might add the PWA to github and maybe let other's actually run with it.
I am so sorry to hear about your diagnosis, it happened in our family and it changed our lives. I will definitely try your app because I am curious if AI can improve diabetes management. Does the app support both mmol/L and mg/dL? Have you had any limitations with iOS? Do you have an Android version in mind? Thanks for your efforts
Pizza is a good example of why not. Slices come in very different sizes, sauces have very different carb content, so do crusts, and toppings.
Edit: for example this pizza(1) is 31g per slice and this pizza(2) is 73g per slice. The difference is very meaningful and the “general idea” given by photo recognition would likely be wrong to the point of dangerous for a diabetic in both cases.
If you’re looking for software that can make a guess simply for the sake of generating a number to write down and not be used in any way, a random number generator would be safer since the risk of output being misconstrued as actual information is much lower.
1 https://www.costcobusinessdelivery.com/kirkland-signature-ca...
Of course if you eat a Neapolitan pizza with not that much of cheese everything changes again. And YMMV, I'm just talking about my experiences.
The people that make the label make the food. They know what they put in it. Because they made it. They wrote down what they put in it for you to read and make decisions off of. The difference is categorical.
Then I calculate how much my serving contains.
Depending on what you eat, what type of diabetes you have and how it’s treated you may have to consider the amount of protein and fat as well (they slow digestion and cause a delayed rise in blood sugar levels). If you have an insulin pump you may want to program a delayed insulin dose to handle that.
Sounds complicated? It is, but only during the first weeks. You quickly learn the carbs content of the food you frequently eat and learn to estimate how much is on your plate. Like, two units for a bun. There are also great nutrition apps out there that help a lot.
No.
I signed up for the free week trial to test out some of the AI features. When I asked it to analyze my week the numbers weren’t very accurate compared to my graphs inside the app.
If you need help troubleshooting I’d be more than happy to help
props on shipping an incredible personal project! if you ever want to geek out about diabetes tech, DM me on X @kamens
scott hanselman would probably also love to chat about your project
https://androidaps.readthedocs.io/en/latest/
This is a life-changing app. It lowered my A1c values from around 8% to 5.5%. What is so special with Android is how easy it is to side-load apps, so you can compile AndroidAPS by yourself and keep using it. In Apple ecosystem you need the developer subscription and you also need to reinstall the app every now and then. There is still the Loop app if any iOS users want to try, but this complication from Apple has just pushed me to Android ecosystem for the past decade already.
https://iaps.readthedocs.io/en/main/
But I like using Loop/LoopKit due its simple interface.
I’m dead I love it.
Having been through the year long build of a similar app for another health condition here’s some thoughts focusing on the GenAI side of things:
- What’s the source of the responses? Is this a RAG system or straight-to-LLM?? We have an ever-growing huge repo of domain expert written content which is the source for our RAG system. It’s far easier to control potential misinformation with a well set up RAG and some tight prompting/guardrails.
- What verification do you have that the responses are correct? We have a group of 50 highly experienced experts in the field who constantly vet the responses to our synthetic set questions and it was eye opening how often our “looks good to me” analysis was off at the beginning.
- The main reason for this questions was that your responses are going to be wrong sometimes, what legal protection do you have? Disclaimers, terms of use at absolute minimum. Who is liable if the answer is wrong? Just presume someone is going to be very upset if your response gives them bunk info.
Either way - great work.
Not the OP but these look like they are Apple's own SwiftUI Charts framework:
I love the app and feel we could have quite some similarities, I too undertook the effort to create a solo-diabetes app for the community, called "Diabetes Cockpit" - i think yours looks much prettier though <3
Would love to connect and have a coffee - tried the contact on the website but its not implemented yet i think ;P
So maybe you just get in touch via schusterlich@gmail.com if you are up for it, i would be happy!
Really good work! Best, Lukas
Here’s a link to an insulin calculator for fat and protein: https://drlogy.com/calculator/warsaw-method
Sources: “The effect of fat and protein was additive, with blood glucose concentrations increasing by 5.4 mmol/L (97.2 mg/dl) at 5 h, the sum of the individual incremental increases for protein and fat” https://diabetesjournals.org/care/article/38/6/1008/37384/Im...
“Meal composition impacts postprandial glucose excursions. Education on the impact of high-fat and high-protein meals and the adjustment of insulin dosing is necessary.” Source: ADA Standards of Care in Diabetes—2024 https://diabetesjournals.org/care/article/47/Supplement_1/S2...
“Match mealtime insulin doses to carbohydrate intake and, additionally, to fat and protein intake.” Source: ADA Standards of Care in Diabetes—2024 https://diabetesjournals.org/care/article/47/Supplement_1/S1...
“Insulin dosing based on carbohydrate plus fat/protein counting reduces the postprandial glucose levels” Source: Pediatric Diabetes Volume 13, Issue 7 p. 540-544 https://pubmed.ncbi.nlm.nih.gov/22765260/
“research and the use of continuous glucose monitoring have shown that other nutritional properties of food, including fat, protein, and glycemic index significantly affect postprandial glucose excursions” https://diabetesjournals.org/care/article/38/6/1008/37384/Im...