How AlignEat reads a menu
A look behind the scan: from photo, through OCR and ingredient resolution, to the safety verdict in your hand.
When you point AlignEat at a menu, four things happen in the next two seconds. Each one is doing real work, and each one can fail in interesting ways.
First, the camera frame is run through an OCR pipeline tuned for restaurant typography — small caps, decorative fonts, multi-column layouts, prices, and inline allergen marks. Off-the-shelf OCR struggles here; ours has been retrained on thousands of real Dubai menus.
Second, every dish title becomes a query into our knowledge graph. "Pad Thai" maps to a canonical dish entity with a fingerprint of likely ingredients — fish sauce, peanuts, egg, soy, gluten — even when the menu doesn't list them.
Third, that ingredient profile is checked against your dietary profile. Allergens get flagged red. Adaptable items ("can be made without peanuts") get flagged amber. Compatible items get the green pill.
Fourth, we surface a single recommendation: eat it, ask the server, or pick something else. The whole flow is designed to give you a verdict before you've put your phone down.
There's a fifth, quieter step too — every scan you submit (anonymised) helps the graph learn. The 1.26 million dishes we currently cover were built one scan at a time, and the next million will be too.