LOADING...

SUCCESS STORIES

  • Propensity
  • Cross Sell/Up Sell
  • Churn
  • Video Surveillance
  • IDP
  • Recruitment
  • Customer Service

PROPENSITY

Problem/Challenges

Predict the propensity of a customer for a specific behavior from historical transactional patterns. Traditional ML methods require significant hand crafted feature engineering.

Innovation

  • 1.Automated feature engineering for transactional financial data.
  • 2.Automated explanations consumable by business owners.
  • 3.Automated cluster naming that assists strategies.

Approach

  • 1.The challenge here is to use historical transactional data and patterns to understand the customer's propensity to buy a new product.
  • 2.Traditional approach, involved segmenting customers based on demographics.
  • 3.A multi layer deep learning network was constructed to analyze the financial transactions and identify the behavioral patterns.
  • 4.The behavioral patterns were then segmented using a clustering techniques, and the clusters were auto named.
  • 5.Based on the behavior, the propensity of the customer to buy financial products was estimated.

Result

Significant lift in GINI score (10-25%) as compared to traditional models in current implementations.

CROSS SELL/UP SELL

Problem/Challenges

Improve upsell/cross-sell conversion rates from historical transaction patterns. Traditional methods require significant hand-crafted feature engineering.

Innovation

  • 1.Automated the process of identifying and segmenting customers financial behavior.
  • 2.Derived micro segments based on customer behavior and insights into micro segments.

Approach

  • 1.The challenge here is to use historical transactional data and patterns to understand the customer’s propensity to buy a new product.
  • 2.Traditional approach involved segmenting customers based on demographics.
  • 3.A multilayer deep learning network was constructed to analyze the financial transactions and identify the behavioral patterns.
  • 4.The behavioral patterns were then segmented using a clustering techniques, and the clusters were auto named.
  • 5.The propensity of the customer to buy financial products was estimated based on behavior.

Result

Significant lift in top 3 Deciles (15-55% spread throughout).

CHURN

Problem/Challenges

Identify customers who are likely to churn by analyzing cross-sectional (demographics, behavioral) and historical transactional data.

Innovation

  • 1.Automated the process of analyzing customer transaction data and derived features with respect to customer behavior.
  • 2.Provided rationale to the client as to why the customer would churn, based on behavioral traits.

Approach

  • 1.The challenge here is to use historical transactional data and patterns to understand the customer’s proclivity to churn.
  • 2.A multi layer deep learning network was constructed to analyze the financial transactions and identify the behavioral patterns.
  • 3.Based on the behavior, the propensity of the customer to churn was estimated using a deep learning network.

Result

89% of the risky customers were bucketed in top 3 Decile for the entire dataset.

VIDEO SURVEILLANCE

Problem/Challenges

Identify suspicious behavior of individuals at ATMs, using real-time CCTV footage. Model needed to run on a Raspberry Pi.

Innovation

  • 1.Developed a DL model that can identify the location of people in a video.
  • 2.Developed a DL model to identify facial key points of a people present within the ATM.
  • 3.Developed light weight models, so that the solution can be implemented on edge device like Raspberry Pi.

Approach

  • 1.Designed a deep learning model to analyze video footage and accurately determine the location of a person within each frame.
  • 2.Designed a deep learning model to analyze people’s face in order to identify key facial points on the face.
  • 3.Designed a pipeline to use the features identified by both deep learning models to determine fraudulence.
  • 4.Ensured that the models were lightweight so that the models could run on a edge device such as Raspberry Pi.

Result

98% accuracy in identifying people and 97% accuracy in identifying facial key points.

IDP

Problem/Challenges

Intelligent Document Processing leading to structured representation. Traditional methods are low on accuracy and require detailed rule specification.

Innovation

  • 1.Ensemble intelligence to analyze scanned documents, allowing higher accuracy than individual algorithms.
  • 2.Automated structure extraction from heterogeneous documents without detailed rule specification.
  • 3.Intelligent confidence computation from multiple signals.

Approach

An AI solution powered by Deep Learning Algorithms, which analyzes scanned documents using an ensemble of models to extract information and structure them in a defined format and has following capabilities:

  • 1.Image processing to remove noise, adjust tilt
  • 2.Converting scanned documents to text by the use of multiple technologies, including custom Deep Learning models
  • 3.Automated layout and format detection (tables, paragraphs)
  • 4.Automated detection of constructs (addresses, invoice items, custom constructs)
  • 5.Entity recognition
  • 6.Knowledge graph creation

Result

Able to get the accuracy and ease of configuration—replacing human involvement in complex workflows.

RECRUITMENT

Problem/Challenges

Requires intelligent analysis of semi-structured resumes to find the best fit for an open position. Standard search techniques still require significant curation.

Innovation

  • 1.Automated the process of analyzing the resume, identifying significant facts and filtering resumes.
  • 2.Generating a knowledge graph.

Approach

  • 1.Designed a multilayered deep learning network to analyze the resumes and extract significant facts from the resume.
  • 2.The facts were then used to construct a knowledge graph about the person.
  • 3.A deep learning model was designed to use the data present within the knowledge graph to predict the propensity of the person getting shortlisted for a job.

Result

Turn around time for finding relevant resumes reduced by 50%

CUSTOMER SERVICE

Problem/Challenges

Comprehend large volume of documents and provide context-sensitive Q&A capabilities. Traditional chatbots are marginally effective.

Innovation

  • 1.Adaptive Query Patterns interprets a wide swathe of questions aimed at structured data sources.

Approach

  • 1.Standard NLP techniques are introduced in the mix--very effective in narrow use cases such as named entity recognition.
  • 2.Reading Comprehension Razorthink Proprietary complex models/ algorithms like Hybrid Deep Learning models, Dynamic Co-attention networks, QA Net that allow us to pinpoint the context of the questions.
  • 3.Adaptive Query Patterns automatically interpret questions aimed at structured data sources.

Result

Customer interaction increased by 24% and response to grievance time reduced by 47%.