One photo, and you know what you are eating.
Kiweat was a nutrition app: point the phone at a plate and get back what is on it, how many calories it has, its nutrients and a food score, with a note on what that dish is good or bad for.
We built the recognition behind it in 2016, in three phases. First on Google Vision, then with our own neural network trained on a catalogue of 500 dishes, with Vision as the fallback and Nutritionix to fill in what the catalogue lacked. Every correction a user made went back as feedback for the next training round. The back end is Python, with Django, TensorFlow and MongoDB.
Photograph
The user takes a picture of the plate.
Recognise
Our network names the dish; Google Vision steps in when it cannot.
Score
Calories, nutrients and a food score come from the nutrition data.
Learn
The user's correction goes back as feedback for retraining.
Working on something like this? Tell us what you need.

