Designing the Future of Software Solutions

Machine Learning & Data Analytics

Organizations are looking to achieve certain business goals and wish to utilize the data collected over the years to predict the behavior of their consumers and related markets. Traditional BI systems aren’t able to meet this need & hence the need for the Machine Learning based models.

Unsupervised Learning models can be employed on business data to do segmentation and clustering analysis, which can help identify parameters influencing specific business outcomes. Identified parameters can help identify specific actions which managers & businesses must take to improve their business performance. Further, using the data corresponding to the identified influencing parameters, Supervised Learning models can be trained which can predict business outcomes.

Our data scientists study the data source and extract the relevant attributes. Our teams then bring in external data that might be relevant to the use case, cleanse it and transform so it is usable for Machine Learning.

KMG offers a “Customer 360” service to its customers, wherein the prospect/client data is enhanced using the above technique.

Unsupervised learning algorithms are used on prepared data to find segments and clusters of data depicting different business outcomes and factors leading to those outcomes

Most use cases also require a further analysis in order to predict events, business outcomes & anomalies that could occur. KMG will train supervised models for such situations, observe results, on the training data, validate on blind data and finally wrap into production-ready web-services. One such recent implementation by KMG was around predicting the incidence of losses/claims given a set of policies.

While Segmentation and clustering analysis results (including factors influencing different business outcomes) are visualized using Tableau or PowerBI, the results related to business outcome predictions are provided as APIs.

Image & Video Analytics

A lot of business data and information is available in the form of scanned documents and images (For example, Submission/Intake forms, Claim Requests, Doctors Notes). These images & documents contain valuable insights. However, because these are images, it makes it impossible to ingest these and consume such data in the business processes.

Image Recognition models, especially Deep Learning models such as Convolutional Neural Network (CNN) or custom feature models built using OpenCV libraries can be used to train classifier models, which can help make sense of the image data.

Images are studied in detail. Image data is pre-processed by applying different image filters (for example Sobel filter) using OpenCV or similar image libraries. Initial analysis of these features is done to look for patterns. Multiple approaches need to be tried to achieve best results.

Next the feature extraction activity is done and the models are built and trained. Based on patterns observed and volume of data, either deep learning models or OpenCV feature extraction followed by normal Machine Learning models (such as Logistic regression, SVM etc.) or a combination are built. Trained models are tested and the results are published – either as APIs or over UI built specifically to visualize this information.

Among the projects done by the KMG team is a complex project involving the extraction of submission data from scanned documents & PDF files (Submission/Intake forms) and bringing them into the Rating Systems / PAS to further processing. RPA bots were deployed in order to push data into the Legacy UI so the data entry validations & edits did not have to be rewritten in the transformation routines.

KMG also has considerable expertise in analyzing live video feeds. One of our clients is into providing Video Surveillance Systems and has tasked us with making the AI-enabled so that the users can search for video clips using a standard text search (for example “Show me all cases where the Cash Register was opened but no Sale was Recorded” OR “Show me all incidents of Slip & Fall during the past three months”.

Natural Language Processing (NLP)

A lot of Business data and information is available in the form of unstructured text. Examples include Comments from the Assessors/Inspectors, Notes Dictated by Doctors, Customer reviews on social media, or information on various public portals.

This information is quite valuable to the business, it is impossible to use in that format. A restaurant owner, for example, would typically need to manually read all the reviews on Yelp in order to understand how the customer perceive their restaurant. On the other hand, if all these reviews could be downloaded, analyzed using sentiment analysis, converted to a form where the system can raise red flags on the negative comments, and identify the causes of concern by doing a simple frequency analysis, all this analysis and review could take maybe five minutes of the manager’s time.

Natural Language Processing techniques (NLP) and Deep Learning models such as Recurrent Neural Network (RNN) and Sequence-to-Sequence/LSTM models can be used to train Classifier models, which can help make sense of the free flow text data.

Text data is studied in detail. Text data is tokenized, lemmatized and other feature extraction activities are done. Feature extracted data is analyzed to look for patterns based on which decision is made on the right approach to meet the requirement

The results are analyzed and, on reaching a certain degree of accuracy, are presented using APIs and/or Dashboards/Portals.

BOTs : Chatbots, RPA Bots (Robotic Process Automation)

Chatbots…

Most organizations have several people-facing processes – including those for handling customers as well as internal employees. These organizations face two issues with human based interfaces. First issue relates to the fact that despite well-defined Standard Operating Procedure (SOP), humans tend to deviate from defined processes leading to degradation of service. Second issue relates to the fact that human agents are not available 24 x 7 to handle requests.

AI driven Chatbots are capable of taking written or spoken requests to handle such processes automatically. They can be trained to handle different types of requests and can be made available 24×7.

The process involves building the BOT using chosen technology platform such as AWS Lex, Google Dialogflow or Microsoft Chatbot Framework. The conversation flow is implemented in the form of intents, utterances and data slots. Back-end database and interfaces are developed. Middle-ware is developed in Python or equivalent and implemented on chosen platform such as AWS Lambda

KMG recently developed a Chatbot that can be used to provide Agents with a Quote. The chatbot is available on the MGA’s Submission Portal as well as an independent Mobile App. The bot is being expanded to include the policy service functionality.