case study

Next-Gen Ordering Experience with NLP


the customer:

A top-5 U.S. restaurant chain seeking to continue delivering world-class customer service at scale looked to machine learning and native language processing (NLP) to automate the process of intaking verbal orders.



To commit to such a major transformation, the company needed a reliable and repeatable NLP solution they could trust to scale across their fleet of stores; and to do that, they needed deep expertise ranging from data science and semantics, to IT infrastructure and engineering.



phData created an opinionated workflow or “assembly line” for getting ML models trained, deployed, updated, and secured — embedding NLP software (SpaCy) into a streamlined Apache Airflow pipeline, with technology-specific configuration best practices and engineering standards built-in.



With a reliable, repeatable, and highly efficient ML assembly line in place, the restaurant is poised to scale their order automation system across the business, shortening wait times for customers and empowering their associates to focus on delivering personable, high-touch service.

Full story: Top-5 U.S. Restaurant Cooks Up Next-Gen Ordering Experience with NLP

A top-5 U.S. restaurant chain ($10+ billion in sales) was growing fast — and so were its wait times at the counter and the drive-thru. They routinely experience massive volume at peak times, during which they need to continue providing the extremely personable, high-touch customer service for which they’re known, even while serving a much higher number of guests.

To keep wait times low and service levels high, the restaurant decided to think outside the box — turning to technology and machine learning (ML). They had already implemented a successful computer vision project (in partnership with phData) to maintain product quality across their hundreds of locations; now, they began to explore how ML might similarly transform the ordering experience.

The idea was to use natural language processing (NLP) to listen to customers’ verbal orders and automatically put them through to the kitchen — automatically transcribing their words and using ML models to interpret the resulting text, translate them into specific order requests according to an index of terms and menu items.

This kind of NLP solution would not only shorten wait times and improve order accuracy, but free up restaurant associates to focus on improving one-to-one service (e.g., by making more eye contact). However, implementing and maintaining such a solution across their hundreds of restaurants — and proving its feasibility within their own business — was going to be a tall order.

A super-sized NLP challenge

From their experience with the computer vision use case, the restaurant chain understood that successfully designing, iterating, deploying, and maintaining an NLP solution of this magnitude would entail major challenges across disciplines ranging from data science and semantics, to IT infrastructure and engineering:
  • No clear path to feasibility — The company had created a promising POC; but manually building out NLP indices and ML models was clearly unsustainable at-scale. Because the initiative would ultimately involve overhauling everything from POC systems to associate training, they needed a sustainable, reliable implementation platform to demonstrate the solution’s viability to the business and make it a reality.
  • A need for end-to-end ML engineering expertise — Their data scientists were experts at designing ML models — not necessarily at engineering how to ingest data into those models, writing great code, or optimally configuring the full stack of specialized technologies involved (from SpaCy to Apache Airflow and Elasticsearch, down to infrastructure platforms) according to the needs of the use case at hand.
  • The complexity of NLP itself — The company had assumed that the total spectrum of phrases used in drive-thru interactions would be relatively narrow. However, accounting for drive-thru-specific jargon, continually evolving menu items and combos, and varying regional speech patterns—such as “may I have a chicken sandwich” vs. “let me get a chicken sandwich” —made it hard. Many of these phrases would be difficult for ML algorithms to learn, but “hard coding” them, by manually adding them to the lexicon, would require manually recreating the index every time.

A proven recipe for ML transformation

To overcome these challenges, the restaurant chain again put their faith in phData; they were confident, based on the success of the computer vision project, that phData’s ML team could bring the required multidisciplinary expertise across NLP, ML, and data engineering.

And the phData ML team delivered. Leveraging the same flexible MLOps Framework they had relied on for the computer vision project, they developed a solution that embeds the NLP software SpaCy into a streamlined Apache Airflow pipeline, with technology-specific configuration best practices and ML engineering standards built-in.

Now the restaurant’s data science team is empowered to configure business rules without worrying about the underlying technology. With a pipeline in place to automate processes like generating SpaCy rules or repopulating Elasticsearch indices, the customer can now self-serve to rapidly iterate through experiments — easily generating indices in order to test and tweak existing rulesets, or to assess brand new ones — until they arrive at the best working solution to accurately interpret and submit customers’ food orders.

Index Generation Process phData

The sweet taste of success

With the new automated NLP workflow developed by phData, the restaurant chain has overhauled their tiny NLP prototype into a reliable, repeatable, and highly efficient ML assembly line with the potential to dramatically improve the ordering experience for their customers.

  • A path to next-gen customer experience — By freeing its associates to focus on connecting with customers and providing personable service, the restaurant can strengthen their customer experience with shorter wait times, higher order accuracy, and a way to create even more of the positive, high-touch customer service interactions that form the backbone of their brand identity.
  • Simplified data pipelines = sophisticated NLP — The phData MLOps Framework makes data ingestion and infrastructure simple through templates and automation, while empowering data science teams to rapidly iterate ML models. This saves developer time (building pipelines in hours vs. days), empowers data teams to do what they do best: iterate experiments and solve business problems.
  • A foundation for sustainable success — With the ability to automate processes such as index generation, it’s now much more feasible to hard-code phrases into NLP models that might be difficult for the algorithm to learn on its own. This is critical for long-term adoption of the solution because it makes it much easier to account for differences in dialect or newly introduced menu items.
  • Baked-in multidisciplinary expertise — The solution, designed by experienced ML engineers with a deep background in both data engineering and data science, is optimally configured at every level for this particular NLP use case. And small tweaks can yield huge results; for example, by disabling unneeded features in SpaCy, we were able to drastically reduce indexing time.
The restaurant chain is now well-equipped to make their order automation plan a reality, and has begun the process of rolling out the solution across the entire enterprise. It’s a massive initiative, requiring significant investment and organizational change, with the potential to transform their business — allowing them to continue delivering their famously exceptional customer experience across an ever-growing fleet of restaurants — and at the heart of it is a proven, repeatable framework for NLP and machine learning.

Ready to learn more about phdata? Let's chat.