We Find Your Best Customers With Machine Learning

CASE STUDY

Brand Engagement

THE BUSINESS CHALLENGE

Attention is scarce, especially on fast-moving platforms like Twitter. Our customer, a technology company, already successfully built a social audience of tens of thousands of followers in their niche, but wanted a more reliable method to generate engaging content without manual human curation.

OUR TECHNICAL SOLUTION

We analyzed our customer’s historical tweet data and those of similar influencers in the same space, identifying topical & linguistic patterns, temporal & seasonal ones, and features unique to the brand. We found that a combination of rule-based automation and a mix of simpler regression techniques were more accurate in predicting the number of retweets and favorites than complex neural network models. Our ensemble models led to an average 30% lift in social engagement.

CASE STUDY

Bot Marketing & Sales

THE BUSINESS CHALLENGE

Consumers now spend more time on messaging platforms than social networks. Our customer, a global CPG company, wanted to diversify their digital engagement with younger audiences to drive product sales. They also wanted to educate their executives on how to optimize chatbot development and marketing.

OUR TECHNICAL SOLUTION

We designed a conversational experience on Facebook Messenger which suggested dynamic product recommendations based on unique user characteristics. The product was ranked #1 in its category in the Facebook Bot Store and enabled the brand to reach a surprising new demographic. Engagement rates with the bot were 5x higher than with comparable email campaigns.

CASE STUDY

Customer Sentiment Analysis

THE BUSINESS CHALLENGE

Every business wants to understand and respond to how their customers feel about their products and services, but survey data is biased and incomplete. Our customer, a media & entertainment company, had previously done studies on customer reviews collected manually, but wanted to extract insights from user-generated content on a myriad of review websites.

OUR TECHNICAL SOLUTION

We first combined disparate data sources into a centralized view and worked with business experts to design an automated analysis solution to better understand product details and contextualize customer reviews. Our insights enabled the customer to rapidly identify customer issues and offer personalized resolutions.

CASE STUDY

Retail Sales Forecasting

THE BUSINESS CHALLENGE

Brick-and-mortar retailers face rising pressures from e-commerce giants like Amazon who reduce foot traffic and cut into already thin margins. In certain regions, rising real estate costs along with reduced consumer purchasing power further decrease profitability and make local stores unprofitable to operate. Accurately forecasting regional sales is critical to assessing which retail locations to prioritize and which ones to close.

OUR TECHNICAL SOLUTION

We worked with a large nationwide retailer with hundreds of stores to forecast regional store sales in order to optimize inventory, staffing, and openings. Building a robust model required pulling in new data sources on economic and demographic trends as well as using a combination of traditional time series analyses (ARIMA, ETS), seasonal models, and a diverse mix of machine learning approaches.

CASE STUDY

Customer Support Automation

THE BUSINESS CHALLENGE

Enabling customers to quickly find and access company and product knowledge is critical to the success of running any business, especially a multibillion dollar public company. Our customer needed to improve the search results and relevance for their consumer-facing FAQs and knowledge base, which they previously powered with keyword search.  

OUR TECHNICAL SOLUTION

We improved the performance dramatically by using machine learning to identify permutations of queries with the same intent in order to map them to the best search results. We designed features using word and sentence level embeddings, classical text mining approaches, and semantic graphs. To achieve a high-performing result, we combined attentional neural network models with traditional machine learning approaches like gradient boosting.

CASE STUDY

Content Moderation

THE BUSINESS CHALLENGE

Keeping the internet safe for everyone is a top priority for our customer, a leading online publisher. However, with over 4M pieces of both editorial and user-generated content and tens of thousands of daily active readers, manual moderation and removal of harmful or illegal content was impossible. Previous solutions were able to flag obvious spam content, but not more nuanced language and media.

OUR TECHNICAL SOLUTION

We mined our customer’s textual data for insights and patterns on language use, identifying key categories of harmful content. We then labeled a large training data set based on our findings using a mix of human-in-the-loop and automated solutions and trained separate models for each context.