Retail

Modern, digital-first consumers are eager to save time and prefer multiple options at their fingertips while shopping online. Shopping at the local store is now a thing of the past. Grocery shopping continues its evolution from being social and interactive to choosing items and checking out with little to no contact with staff. Before the COVID-19 outbreak began, global consumers quickly adopted contactless payments and automated convenience stores, such as Amazon Go. Retail giants have introduced contactless and mobile checkout options, which have become increasingly important as the pandemic rages on. In a highly competitive environment, retailers must cater to customers who value flexibility and speed, making traditional, manual, and rule-based processes insufficient. As today’s retail environment evolves, retailers continue using AI for a variety of functions. They can gain market share and regain customer loyalty through superior customer experiences and operations, track what products shoppers pick up, automate billing, and eliminate the need for checkout lines.  

By combining the power of our flagship framework — xpresso.ai — with a decade of wholesome experience, we have ensured AI projects are delivered in a timely and robust manner, and organizations achieve a successful AI transformation journey with our systematic approach towards enterprise AI and MLOps solutions.  

Our approach to an enterprise MLOps journey begins with a robust data intake and elaborate analysis to find problem areas.  The second part includes applying an engineering mindset, identify enablers required for AI initiatives to be successful at enterprise scale, and preparing cognitive models. Finally, these models are deployed and managed throughout the lifecycle of the solution. Throughout the lifecycle, xpresso.ai pre-built frameworks are used to push the required transformations ahead. Coupled with a robust uptime for applications we use and near-zero chances of interruption, this means unwavering support all the way for customers. So, what would have typically taken months to deliver can be developed and deployed in weeks, and most importantly — at a fraction of the cost.

Challenges

Lost opportunities due to slow adoption: Despite having the finances and capability to eliminate brick-and-mortar checkouts, retailers are reluctant to include contactless payments as part of the customer experience; most of them are yet to adopt it. J C Penney has suspended contactless payments to comply with the mandatory support for EMV contactless chips without giving a date. ​Although 7-Eleven, Staples, and CVS have adopted contactless payments, it hasn’t gained as much traction in the US as in other countries.

One size doesn’t fit all: Walmart tried out a cashier-less system based on scanning barcodes for about six months in more than 100 stores but discontinued it in April 2018. The technology proved impractical for pricing produce and other items that had to be taken to a cashier to be weighed, causing delays. Due to the contactless limit, not every business transaction can leverage a contactless feature (although it remains attractive for low-value payments).

Security concerns: Tap to pay cards don’t offer the additional security of biometrics and two-factor authentication of mobile payments. They also don’t provide the slick financial management tools of mobile wallet payment apps. Contactless payment limits might be bypassed through a ‘man in the middle attack.’ In fact, a Visa vulnerability allowed a bypass on the card limits with five major UK banks. 

Discriminatory policy and privacy: The cashier-less store concept has faced some challenges due to a backlash against contactless systems for discriminating against customers without bank accounts and undermining privacy and data security. The Federal Deposit Insurance Corporation estimated that 20.5 million people (6.5% of US households) didn’t have a checking or savings account; subsequently, they weren’t eligible for contactless payment accounts.

xpresso.ai – Cognitive Solutions

• Increased foot traffic at stores means longer lines at the checkout counters, higher waiting time for customers, and a rise in negative customer experience. A large Hong Kong-based retail chain leveraged our expertise to overcome such challenges.
• Our cognitive expertise and xpresso.ai’s MLOps environment helped reduced the checkout time by a whopping 75% — from ~6 seconds to ~1.5 seconds, across 350 outlets.

Solution Approach

Use Case Discovery

We actively engaged with our client to capture the business requirements while observing the problem. We identified the relevant datasets and formulated a case-based approach that would solve the business problem or improve/predict actionable insights to mitigate the problem. 

We assisted in setting up the required infrastructure – xpressso.ai framework provides out-of-the-box development platforms, and all functions can be accessed using a Jupyter Notebook — ensuring zero-delay and plug-and-play availability of high-end hardware. Development images configured based on pre-defined templates can be installed on-premises or in a development VM within the xpresso.ai infrastructure. This enables authentication using LDAP, seamless project setup using Bitbucket, Jenkins, and Docker (ensuring build and deployment without software compatibility issues).  The project was started seamlessly with the relevant environments, which are subsequently created automatically.

The framework made available by xpresso.ai leverages the latest ML and DL tools while preparing models and includes Pachyderm-based data versioning, deployment using a Kubernetes orchestration system, Kubeflow and Spark-based ML and DL build and deployment, Istio-based service mesh enabled microservice architecture, and ELK based monitoring capability, contributing to reduction in latency time.

We also set up the required hardware — camera/mobile camera positioned on top of a tripod at an angle to have a 2.5′ X 2.5′ white tabletop surface.​

Data Engineering

xpresso.ai’s MLOps framework has different data adapters that are available through a common catalog of services which simplifies interoperability and scalability concerns, enables APIs, and abstracts all the technical complexities from the service consumer. This allows establishing high-end Alluxio and Presto-based rapid, inexpensive data connectivity and data collection from diverse sources (available in structured, unstructured, and streaming formats) coming in at a high velocity and in huge volumes. 

All the data sources are funneled into the data storage layer after proper validation and cleansing. The storage landscape with different storage types and extreme flexibility is built-in to manipulate, filter, select, and co-relate different data formats.

We obtained images and had SKU/instance-specific bounding box annotations on the captured scene images from cameras and passed these through a DNN-based algorithm that we had created to detect the correct SKU instances from a scene.

Infrastructure and MLOps Automation

The details collected, project code, data preparation workflows, and models can be easily versioned in a repository (Bitbucket, Git, etc.), and data sets can be versioned through on-premise /cloud storage. These can be added as exploratory variables by using two excellent features of xpresso.ai: 

1. Data Connectivity Marketplace libraries
2. Data Versioning

The attributes obtained are used for categorization (employing Pachyderm-based data versioning) and then performing univariate, bi-variate, and Bag of Words analysis — for both structured and unstructured datasets through xpresso Exploratory Data Analysis (Data and Statistical Analysis).  Different datasets and their different versions can be easily controlled and stored into xpresso Data Model (XDM)-enabled data store that enabled easy retrieval and storage of datasets/ files into internal XDM. 

xpresso.ai MLOps automation allows creating pipelines, train with as much data and as accurately as possible, fastest time to inference, with the ability to rapidly retrain. The xpresso Data Pipeline Management (Rapid Model Training and Experimentation) uses Kubeflow-enabled pipelines. Thus, multiple experiments using different models and datasets can be created, tested, paused, and restarted to gain better insight. xpresso.ai’s MLOps environment helped create object detection using Tensorflow for ten classes, fast R-CNN with ResNet101-based model and achieve a 93% accuracy (mean precision) while creating a Computer Vision object detection model. The best performing inference model generated was published on cloud/on-prem/device. The JSON response containing a list of SKU/barcodes was used to communicate with the billing software used at the stores. Afterward, the barcode/SKU info was used to process the bills.