This publish was co-authored by Arun Gupta, the Director of Enterprise Intelligence at Prodege, LLC.
Prodege is a data-driven advertising and marketing and shopper insights platform comprised of shopper manufacturers—Swagbucks, MyPoints, Tada, ySense, InboxDollars, InboxPounds, DailyRewards, PollFish, and Upromise—together with a complementary suite of enterprise options for entrepreneurs and researchers. Prodege has 120 million customers and has paid $2.1 billion in rewards since 2005. In 2021, Prodege launched Magic Receipts, a brand new means for its customers to earn money again and redeem present playing cards, simply by procuring in-store at their favourite retailers, and importing a receipt.
Remaining on the chopping fringe of buyer satisfaction requires fixed focus and innovation.
Constructing an information science crew from scratch is a good funding, however takes time, and sometimes there are alternatives to create fast enterprise influence with AWS AI providers. In accordance with Gartner, by the top of 2024, 75% of enterprises will shift from piloting to operationalizing AI. With the attain of AI and machine studying (ML) rising, groups have to deal with the right way to create a low-cost, high-impact answer that may be simply adopted by a company.
On this publish, we share how Prodege improved their buyer expertise by infusing AI and ML into its enterprise. Prodege wished to discover a option to reward its clients sooner after importing their receipts. They didn’t have an automatic option to visually examine the receipts for anomalies earlier than issuing rebates. As a result of the quantity of receipts was within the tens of hundreds per week, the guide means of figuring out anomalies wasn’t scalable.
Utilizing Amazon Rekognition Customized Labels, Prodege rewarded their clients 5 occasions sooner after importing receipts, elevated the proper classification of anomalous receipts from 70% to 99%, and saved $1.5 million in annual human evaluation prices.
The problem: Detecting anomalies in receipts shortly and precisely at scale
Prodege’s dedication to top-tier buyer expertise required a rise within the velocity at which clients obtain rewards for its massively in style Magic Receipts product. To try this, Prodege wanted to detect receipt anomalies sooner. Prodege investigated constructing their very own deep studying fashions utilizing Keras. This answer was promising in the long run, however couldn’t be applied at Prodege’s desired velocity for the next causes:
- Required a big dataset – Prodege realized the variety of photos they would wish for coaching the mannequin could be within the tens of hundreds, and they’d additionally want heavy compute energy with GPUs to coach the mannequin.
- Time consuming and expensive – Prodege had lots of of human-labeled legitimate and anomalous receipts, and the anomalies had been all visible. Including extra labeled photos created operational bills and will solely perform throughout regular enterprise hours.
- Required customized code and excessive upkeep – Prodege must develop customized code to coach and deploy the customized mannequin and preserve its lifecycle.
Overview of answer: Rekognition Customized Labels
Prodege labored with the AWS account crew to first determine the enterprise use case of having the ability to effectively course of receipts in an automatic means in order that their enterprise was solely issuing rebates to legitimate receipts. The Prodege information science crew wished an answer that required a small dataset to get began, might create fast enterprise influence, and required minimal code and low upkeep.
Primarily based on these inputs, the account crew recognized Rekognition Customized Labels as a possible answer to coach a mannequin to determine which receipts are legitimate and which of them have anomalies. Rekognition Customized Labels supplies a pc imaginative and prescient AI functionality with a visible interface to routinely practice and deploy fashions with as few as a few hundred photos of uploaded labeled information.
Step one was to coach a mannequin utilizing the labeled receipts from Prodege. The receipts had been categorized into two labels: legitimate and anomalous. Roughly 100 receipts of every type had been rigorously chosen by the Prodege enterprise crew, who had data of the anomalies. The important thing to a superb mannequin in Rekognition Customized Labels is having correct coaching information. The subsequent step was to arrange coaching of the mannequin with a number of clicks on the Rekognition Customized Labels console. The F1 rating, which is used to gauge the accuracy and high quality of the mannequin, got here in at 97%. This inspired Prodege to do some extra testing of their sandbox and use the skilled mannequin to deduce if new receipts had been legitimate or had anomalies. Organising inference with Rekognition Customized Labels is a simple one-click course of, and it supplies pattern code to arrange programmatic inference as properly.
Inspired by the accuracy of the mannequin, Prodege arrange a pilot batch inference pipeline. The pipeline would begin the mannequin, run lots of of receipts towards the mannequin, retailer the outcomes, after which shut down the mannequin each week. The compliance crew would then consider the receipts to verify for accuracy. The accuracy remained as excessive for the pilot because it was in the course of the preliminary testing. The Prodege crew additionally arrange a pipeline to coach new receipts with the intention to preserve and enhance the accuracy of the mannequin.
Lastly, the Prodege enterprise intelligence crew labored with the applying crew and help from the AWS account and product crew to arrange an inference endpoint that may work with their utility to foretell the validity of uploaded receipts in actual time and supply its customers a best-in-class shopper rewards expertise. The answer is highlighted within the following determine. Primarily based on the prediction and confidence rating from Rekognition Customized Labels, the Prodege enterprise intelligence crew utilized enterprise logic to both have it processed or undergo extra scrutiny. By introducing a human within the loop, Prodege is ready to monitor the standard of the predictions and retrain the mannequin as wanted.
Prodege Anomaly Detection Structure
Outcomes
With Rekognition Customized Labels, Prodege elevated the proper classification of anomalous receipts from 70% to 99% and saved $1.5 million in annual human evaluation prices. This allowed Prodege to reward its clients 5 occasions sooner after importing their receipts. The perfect a part of Rekognition Customized Labels was that it was simple to arrange and required solely a small set of pre-classified photos to coach the ML mannequin for prime confidence picture detection (roughly 200 photos vs. 50,000 required to coach a mannequin from scratch). The mannequin’s endpoints could possibly be simply accessed utilizing the API. Rekognition Customized Labels has been a particularly efficient answer for Prodege to allow the sleek functioning of their validated receipt scanning product, and helped Prodege save plenty of time and sources performing guide detection.
Conclusion
Remaining on the chopping fringe of buyer satisfaction requires fixed focus and innovation, and is a strategic aim for companies right this moment. AWS pc imaginative and prescient providers allowed Prodege to create fast enterprise influence with a low-cost and low-code answer. In partnership with AWS, Prodege continues to innovate and stay on the chopping fringe of buyer satisfaction. You will get began right this moment with Rekognition Customized Labels and enhance what you are promoting outcomes.
In regards to the Authors
Arun Gupta is the Director of Enterprise Intelligence at Prodege LLC. He’s obsessed with making use of Machine Studying applied sciences to offer efficient options throughout numerous enterprise issues.
Prashanth Ganapathy is a Senior Options Architect within the Small Medium Enterprise (SMB) section at AWS. He enjoys studying about AWS AI/ML providers and serving to clients meet their enterprise outcomes by constructing options for them. Outdoors of labor, Prashanth enjoys pictures, journey, and attempting out totally different cuisines.
Amit Gupta is an AI Companies Options Architect at AWS. He’s obsessed with enabling clients with well-architected machine studying options at scale.
Nick 
Ramos is a Senior Account Supervisor with AWS. He’s obsessed with serving to clients remedy their most advanced enterprise challenges, infusing AI/ML into clients’ companies, and assist clients develop top-line income.
