The Ass1stant D — A Meal Assistant for T1D Teens

by a biomedical AI researcher & first-time LLM builder

Kaggle notebook

Github repo

Where it all started

This project began with a phone call. My 14-year-old cousin had just been diagnosed with type 1 diabetes. She was overwhelmed, and so was I. I wanted to help, but despite my two years in biomedical AI, I had never worked with language models. However, something about GenAI sparked my curiosity, and this time, it was personal.

The Gen AI Intensive Course came at the perfect time! With the expertise I gained in LLMs and agents, I set out to build something that could help her—not years from now, but right away. Something approachable and practical. A small AI companion to help her (and others like her) make smarter food choices, without sounding like a textbook or a glucose monitor.

The Ass1stant D

The Ass1stant D is a meal-planning assistant designed for people living with T1D. It doesn’t just suggest “healthy meals”—it considers glycemic load, fiber, insulin dosing, and the kind of questions a teenager might ask. It aims to talk like a friend who understands their biology.

Technically, it’s an agent-powered RAG pipeline backed by nutritional advice and custom logic for T1D-specific analysis. But behind the tech, it’s simply trying to answer: “What can I eat that won’t spike my glucose later?”

What it does

• Takes in a user profile: age, weight, insulin plan, etc.
• Parses food input (from ingredients or meals) and scores it for glycemic impact
• Suggests alternatives or confirms good choices
• Explains the reasoning in plain language

GenAI techniques used

1. Agents: A multi-turn agent that uses reasoning and tools to walk through carb scoring, substitutions, and explanations.
2. RAG: Provides grounded answers using vector search over USDA data and hand-picked diabetes-safe recipes.
3. Function Calling: Offloads nutrition analysis, insulin ratio math, and glycemic estimation to Python functions.

What I learned

This was my first real dive into APIs, LangChain, Gemini, and prompt engineering. I had zero experience building agents before this project. It wasn’t smooth, but it worked. Now I know that with the right model, a good prompt, and a clear goal, GenAI can be genuinely useful—even in healthcare.

Limitations

• This is just an MVP. Right now, document chunks are manually selected and injected into the vector store.
• It's not a medical device and cannot replace a clinician or registered dietitian.
• Ingredient parsing and matching are simplistic for now.
• It assumes accurate and honest input from the user.

What’s next

I’m already thinking about what a final version might look like:

• Add a carb estimator module that leverages object recognition and is based on medical checked tables for counting.
• Replace the hand-curated chunks with a full medical knowledge base—parsed, embedded, and queried with LangChain.
• Integrate CGM APIs for feedback loops based on real glucose response.
• Expose the assistant via a chat UI (and maybe voice) to make it more accessible.
• Add a caregiver view for shared decision-making.

Final thoughts

This wasn’t just an experiment in GenAI—it was a way to channel concern into something useful. If even one kid with T1D finds it helpful, that’s more than enough to make this worth it.

— G.