ai-food-analyzer

AI-powered food photo analyzer with nutrition analysis

Comment

Inspiration

Most nutrition apps today are too complicated — they require manual logging of ingredients, portions, and brands.
I wanted something simple and instant: snap one photo and get a quick idea of how healthy the food is.

I was inspired by research from top universities:

  • Tufts University – Food Compass system
  • University of Oxford – Nutrient Profiling Model (NPM)

These frameworks prove that food health can be scored systematically. My goal was to condense this science into a single 0–10 health score that anyone can understand at a glance.

What It Does

The app is a mobile-friendly web tool with one mode: Ingredient Analyzer.

  • Upload one food photo
  • AI predicts calories, sugar, and fat per 100g
  • A 0–10 health score (1 decimal place) is calculated
  • The system provides a short nutritionist-style tip, such as:
    • “High sugar — limit intake.”
    • “Balanced nutrients — good choice.”

This makes food evaluation fast, clear, and accessible, even for users without nutrition knowledge.

Challenges We Ran Into

  • Image ambiguity: sauces, mixed dishes, or hidden ingredients reduce accuracy.
  • Portion vs per-100g confusion: science prefers per-100g scoring, but users think in servings.
  • Data scarcity: hard to benchmark predictions without a large labeled dataset.
  • On-device constraints: ensuring models run under 5 seconds on mid-range phones.

Accomplishments That We're Proud Of

  • Built a working MVP that analyzes one photo and gives a clear health score.
  • Achieved fast on-device inference (~3–5 seconds per photo).
  • Added nutritionist-style advice, making the score more understandable.
  • Designed a minimal, mobile-first UI for frictionless use.

What We Learned

  • Scientific scoring systems (Food Compass, NPM, HSR) can be adapted into a simple user-facing number.
  • Explainability matters — users trust the score more when paired with a short advice line.
  • Mobile-first design forces efficient preprocessing and model optimization.
  • Accuracy is important, but speed and clarity are equally critical for user adoption.

What's Next for AI-Food-Analyzer

  • Add more nutrient predictions: sat fat, sodium, protein, fiber.
  • Calibrate against Food Compass or Oxford NPM scores with larger datasets.
  • Introduce confidence ranges (e.g., 6.2–6.8 instead of just 6.5) when predictions are uncertain.
  • Explore portion-size estimation using hand/utensil scale cues.
  • Long-term: expand into multi-photo support and recipe generation mode.
Share this project:

Updates