Calorie Scan Accuracy

Calorie scan accuracy measures how closely AGNT's food photo analysis matches professional nutritionist estimates. When a user sends a food photo, the system uses vision-capable LLM inference to identify food items and portions, then cross-references against the Nutritionix database for USDA-backed nutritional data. The '92% within 10%' means that for 92 out of 100 scans, the total calorie estimate falls within plus or minus 10% of what a certified nutritionist would estimate for the same photo.

The accuracy varies by food type. Simple single-item foods (a banana, a slice of pizza) hit 97%+ accuracy because the Nutritionix database has precise per-item data. Mixed plates and Southeast Asian dishes (nasi campur, lawar) are harder — accuracy drops to ~85% because portion estimation from photos is inherently imprecise for mixed dishes, and some local dishes have limited coverage in Western-centric nutrition databases.

We're improving mixed-dish accuracy through two approaches: fine-tuning portion estimation prompts with SEA-specific food photography, and building a local nutrition database supplement with calorie data for 500+ Balinese and Indonesian dishes sourced from Indonesian Ministry of Health publications.