NutriSnap

Health Dashboard
Fridge Snap
Recipes recommandations
Personal Nutrition Coach

Inspiration

What do we eat tonight... UberEats ? It is often tiresome to find ideas of recipes, and even more of good healthy recipes with limited ingredients. Thus the idea to scan the ingredients of your refrigerator and have a proposition of healthy recipes with those ingredients. Food is indeed one of the main factors to stay healthy.

What it does

That is why after the scanning, we offer to browse recipies along with their "nutriscore" and percent daily values of sugar, fat, protein... We then keep in memory (database) the calories and nutrional values of the recipe and compare it to personalized goals (depending on weight, age, genre, height...) as well as taking into account the amount of sport practiced. We also have a chatbot, a personalized coacg, who knows all our stats and can help us with all our health related questions.

How we built it

We coded in Python and used Streamlit for the front end, sqlite for the backend. We fine tuned Yolov11 on 30 common ingredients to be able to do box detection. We found and preprocessed a large recipe dataset, ranked its recipes by relevance and linked them to the official website. We used OpenAI API for the Chatbot, which we fed the data of the user including sports activities from the Garmin API, to generate a personalized and optimized training program.

Challenges we ran into

First AI models not very effective (CLIP)
Building a coherent database taking every personnalized data into account
Finding interesting recipes dataset and preprocessing it into something usable (dozen of thousands of ingredients to summarize into 30) and rankable (nutri-score)

Accomplishments that we're proud of

We took care of building a highly functional application with a login for personalized advices for the long term. We used a state of the art computer vision model, fine tuned efficiently on what we need instead of using simply CLIP which add lesser results.