MSc Data Science alumni develop an open-source AI model helping restaurants find patterns in customer reviews
- Alumni
The MSc Data Science programme provided exactly that. It pushed them to think rigorously about data, about models, and about what it actually means to solve a problem rather than just demonstrate one.
While studying and after graduating, the pair worked part-time in front-of-house hospitality in London to support themselves. It was in that world that they noticed something. Restaurants and pubs collect huge amounts of customer feedback through online reviews, but almost none of them have the tools to make sense of it systematically. The language of food is specific: a review that mentions a dish being over-seasoned or a technique being inconsistent carries a real operational signal. But general AI language tools are not built to read it that way.
After graduating from Kingston, Akshatha and Siddhant spent time working in the AI industry, gaining a further grounding in how real-world machine learning systems are built and deployed. Eventually making the decision to back themselves, the duo co-founded Morze Tech, where they've been building their own hospitality technology product. But alongside that, and entirely separately from the start-up work, Akshatha and Siddhant kept coming back to the gap they had spotted in the restaurant sector. No publicly-available Natural Language Processing (NLP) model existed that was trained specifically on culinary language in a UK context. So they built one.
At Kingston University, I learned that good data science starts long before the modelling, it starts with understanding the problem deeply enough to know whether you're solving the right one. The Machine Learning and Artificial Intelligence module gave me both the tools and the confidence to do that. That foundation stayed with me after graduating and CulinaryNER was the result.
CulinaryNER is an open-source named entity recognition model trained on over 8,000 restaurant reviews spanning five cuisine types. Of these, 2,500 were manually curated from UK-based restaurants and annotated by hand. Working in their free time, on weekends and after work, they believed it was worth doing regardless of whether it had a commercial application. Akshatha led the data engineering and annotation work. Siddhant focused on model development and training. The model was released freely on GitHub under a Massachusetts Institute of Technology (MIT) licence so that anyone, researchers, developers, or small hospitality businesses, could use it without any cost.
The model was put to an unplanned but meaningful test when a customer engagement specialist at a major London pub group agreed to review its outputs. Working from publicly available review data alone, CulinaryNER had independently flagged a pattern of declining sentiment around a specific menu item. When he checked it against internal records, the same issue had already been identified and resolved through a specific product change from the supplier. We had arrived at the same conclusion through review analysis alone, with no access to any internal information, and done so considerably faster than the internal process had. For a smaller venue without dedicated operations teams, that kind of early signal from customer feedback could make a real difference.
A restaurant can serve hundreds of covers a week and never surface the patterns sitting in their own customer reviews. Seeing how much useful feedback gets lost in the industry, the pair's goal was to build something that could change that, and make it available to everyone rather than just the businesses with budgets for enterprise software. The industry presented the problem. Kingston University provided the tools to solve it.
CulinaryNER is freely available.
Kingston University helped me identify real-world problems and think about practical solutions with genuine impact. The Machine Learning and Artificial Intelligence module kept the emphasis on building things that actually work outside a classroom. That mindset shaped every decision I made on CulinaryNER because I believed the hospitality sector deserved better tools, not because it was part of any commercial brief.
