Skip to content
WebSurge

BiteBrunch: from a photo of a dish to a calorie range

A food photo is a rough source. Letting a model invent calorie numbers would look precise and be wrong, so the model never computes them.

· 2 min read · By WebSurge Technologies

At a glance

Summary

StatusIn development. The recipe community has 488 recipes published. The AI scan ships switched off until the owner enables it.
TypeRecipe community and cooking tools with a photo-based calorie scan.
StackNext.js 16, PostgreSQL 18, the Claude API, sharp for image processing and Zod for validation.
Facts as of16 November 2026, from the project’s plan and decision log.

The problem

A photo is not a measurement

The scan lets a member photograph a dish and see what it contains. But a single photo hides portion size, oil and sauces, and photo-based calorie estimates are typically 20 to 30 percent off. A model that returns one confident number would mislead people, and a model with an open API key is also an open bill.

Our approach

Split the job between the model and the server

  1. Upload. The member uploads a photo.
  2. Prepare. The server resizes the image and removes its EXIF data.
  3. Read. A vision model returns JSON that names the components of the dish and estimates their portions in grams.
  4. Compute. The server computes calories from an ingredient dictionary based on USDA data.
  5. Show. The member sees a range, labeled as an estimate, not a single number.

Because the server owns the arithmetic, the numbers can be tested and explained, and the model cannot drift outside its lane.

Safety and cost

Limits that an admin can change

  • Defaults of 5 scans per member per day and a monthly spending cap of 10 US dollars.
  • Both limits are admin settings, so the owner can change them without a release.
  • A kill switch, and the scan ships switched off until the owner enables it.
  • Tests run against a fake Anthropic server, so they never spend money.

Honest status

What is not proven yet

BiteBrunch is still in development. The limits above are design defaults from the project documents, not outcome data, and we do not claim accuracy figures for the scan. The 488 recipes are published content, not traffic or sales.

Start a project

Have a platform or an AI idea that has to work in production?

Tell us what you are building. The first consultation is free, and you get a short technical read-out and a plan.