F100 Media Broadcaster – ML Engine based next best recommendation​

Business problem

Client wanted to drive better customer experience and up-sell/cross-sell conversion over the huge volume of interactions between customers and contact center  agents. Client call volume: 25M+ incoming calls a year from 15M+ customers​.

Fractal solution

• Fractal architected a customer-first approach in the form of top 3 product recommendations from 30 product lines for the entire customer base on the cloud platform.

• Built AI-driven recommendation systems, post customer 360 development with the help of 25+ data sources. Cloud architecture to incubate harmonization solution of input data pipelines as well as the modeling assets.

• AI model outputs are served via API that services call agent screens to surface top-3 product recommendations.

• The contextual layer comprising of customer insights to provide ‘talking pointers’ to call agents vis-à-vis rationale of respective product relevance

Business outcomes & impact delivered

Big data sources like viewing and clickstream provided a significant boost in the understanding of customers’ product affinity. Global top drivers from models, when contextualized to the customer journey, provided call agents with a relevant conversation context.

First sophisticated cloud tech stack to support enormous computation of weekly acceptance likelihood of 30 products for 15M+ base​

Five thousand call agents and retail shop agents use this tool regularly. Swift adoption and excellent agent feedback; high % of agents showed an uplift in sales in 1st week itself.

Improved accuracy from benchmark models by an average of 18%. The annual incremental benefit of $10M+ is based on higher upgrade sales volumes, better sales mix, better save rate, etc.

Google Cloud services​

1. Big Query

2. Compute Engine

3. Cloud Storage

4. Cloud Composer

5. Cloud SQL

6. AI Platform Notebook

Value addition / accelerators​

• Scalable TensorFlow based AI architecture to host 3 sets of models – Linear, DNN & Boosted tree classifiers across 30+ products

• Average models build to productionization time for new products significantly reduced from 4-6 weeks to 1-2 days using automated, scalable AIML architecture

• Highly modular cloud-based architecture to orchestrate streamlined data processing and AI-powered modeling pipelines (including model training, integration, and serving) at scale