Nutritionist LLM

Overview

Food-logging apps fail because logging is tedious: searching for each item, guessing portion sizes, typing it all in. Nutritionist’s Corner skips that. Upload photos of your meals and a multi-modal LLM acts as a nutritionist, identifying what’s on the plate and assessing how healthy it is.

You can upload several photos at once, for example breakfast, lunch, a snack and dinner. The app combines them into one report for the whole day, rather than one report per photo.

What the report includes

1. Food items: every item identified across all the photos.
2. Nutrition per item: calories, fat (total, saturated and unsaturated), energy, carbohydrates, protein, and other key components such as fibre, sugar and sodium.
3. Totals: total calories and the macronutrient split for everything eaten.
4. Health rating: a score from 1 to 10 based on nutrient balance, calorie density and essential vitamins and minerals, with an explanation of the score.
5. Recommendations: specific swaps or additions to improve the meals, plus an estimate of the revised health score if you follow them.

How it works

• One model call. The uploaded images are sent to Gemini 2.0 Flash together with a detailed nutritionist prompt. Food recognition, nutrition estimates and advice all happen in that single call, instead of chaining an image classifier, a nutrition database lookup and a separate text model.
• The prompt does the structuring. It lists the exact sections the report must contain and the criteria the model should be judged on: identification accuracy, precise calculations, useful recommendations and clarity. It also tells the model to consider all images and produce one combined report.
• Simple upload handling. JPG and PNG images are saved to a working folder that is cleared on each run, shown back as thumbnails, and read as Part objects with their MIME type for the API.
• Timing. Each analysis shows how long it took.

Limitations

The figures are estimates from a photo. Portion sizes, hidden ingredients such as oil or sugar, and cooking methods are hard to judge visually. Treat the output as a quick, informed guide, not as medical or dietary advice.

Running it

pdm install
echo "GOOGLE_API_KEY=<your key>" > .env
pdm run streamlit run app.py