From a photograph to something useful

The premise was intentionally simple: upload a photo of food and get back a practical list of likely ingredients.

I built the experience with Next.js and TypeScript, using GPT-4 Vision for image understanding and Vercel for deployment. The project reached 92% ingredient accuracy in testing, but its real value was as a compact interface experiment, taking a probabilistic model response and presenting it in a form someone could quickly understand and use.

It was small enough to move from idea to a working product without hiding behind infrastructure, which made the quality of the interaction itself the main engineering constraint.