Every restaurant has a shopping list. The problem is, most food suppliers can only see their own line item.
Say you're a distributor selling produce to a restaurant. You know how much produce that restaurant buys from you. What you don't know is whether they're buying twice as much meat from a competitor, or whether their menu is quietly begging for a dairy order you're not even in the room for. That blind spot isn't just an inconvenience — it's a missed opportunity to grow the account, one that shows up on distributors' balance sheets across the industry.
It turns out the answer to that blind spot was sitting in plain sight the whole time: the menu itself.
Restaurant menus are dense with information — dish names, ingredients, quantities, prices — but almost none of it is stored anywhere useful. Useful data likely lives in PDFs, photographs, and hand-designed layouts rather than in a centralized location.
A multi-billion-dollar American foodservice distributor came to us with a scale problem of this sort: how do you turn hundreds of thousands of restaurant menus into something a sales team can act on? Not "figure out one restaurant." Figure out all of them, fast enough to make a real dent in revenue.
Synaptiq treated this issue with object detection. We built a machine vision model that can draw a box around each individual dish on a menu, then pull the name, price, and ingredients out of that box.
Simple in concept, a lot more difficult in practice. Menus don't follow a universal template — every restaurant has a distinct design with its own fonts, layout, often as an image rather than clean text. Feed that into standard optical character recognition (OCR), and you'll get plenty of mistakes.
So we gave the system some context to reason with. Known ingredient lists, typical price ranges, and cuisine type all became signals the model could use to catch and correct its own errors — a menu item priced at $4.99 for a steak dinner is probably a misread, not a bargain. We also kept every raw menu on file rather than throwing it away after extraction, so the models could be rerun and improved later without re-scraping the entire web from scratch.
The system went into production. Our client can now build a picture of what a restaurant is actually buying — and, just as importantly, what it should be buying but isn't — turning a guessing game into a targeted sales strategy. It's the kind of proactive, personalized service that "wows" existing customers while surfacing entirely new restaurants that look like strong prospects.
Our Chief Data Scientist, Dr. Tim Oates, recently revisited this object detection problem and ran the same menu through a general-purpose LLM to see how the LLM stacked up against our custom-built model.
Asked broadly to pull dishes from a menu link, the model answered confidently — and wrongly. The LLM hallucinated, listing a dish that never appeared on the actual menu at all. But when Dr. Oates narrowed the ask, dropping the actual menu text in front of the model and asking specific questions, the results were strong: correct prices, correct ingredients, sensible answers to natural follow-up questions.
The takeaway is one every team building with AI should sit with: general-purpose LLMs can get you most of the way there, remarkably fast. But a broad, unconstrained prompt is an invitation for a confident wrong answer. The output still needs a human — or a well-built system — checking its work at all times.
This project is a good stand-in for a pattern we see across industries: the data you need to grow already exists, but it's trapped in a format built for humans to glance at, not for systems to learn from. Menus, contracts, inspection reports, medical charts — the shape of the problem repeats. The food and beverage industry has great AI potential, but AI applications extend far into many other industries as well. Machine vision and OCR, paired with just enough domain knowledge to keep the system honest, often help companies unlock the value of their data.
Growing an account isn't always about selling harder. It's about seeing more clearly what a customer already needs — and menus, it turns out, are more than happy to tell you.
Photo by Jakob Owens on Unsplash
Synaptiq is an AI and data science consultancy based in Portland, Oregon. We collaborate with our clients to develop human-centered products and solutions. We uphold a strong commitment to ethics and innovation.
Contact us if you have a problem to solve, a process to refine, or a question to ask.
You can learn more about our story through our past projects, blog, or podcast.