Predict anomalous vendor transactions in procurement work-flow, and hence strengthen client audit team.

• Challenges: Several possible cross-functional reasons for anomalous transactions. Possible reasons can come from accountings, country-wise regulations, BU practices etc. We need to reveal all those possibilities from volume of data.
• EDA, Feature Engineering & Data Transformation
• POC: Research on suitable algorithms (Isolation Forest, LOF, Elliptical Envelope)
• Educating client on ML models
• Results: Customer Penetration
  • Received wide acceptance from senior management from client
  • Won project extension
  • Cross selling / Up selling: Won another project from the client

Predict demand fulfillment likelihood & prioritize the same, and help the client to increase business efficiency.

• Challenge: Imbalanced dataset (extremely low conversion rate), unclean & unstructured textual data
• EDA, Data cleaning using NLP & Data Transformation, Oversampling (SMOTE)
• POC: Research on algorithms (Supervised: Tree-based, SVM, Naïve-Bayes etc)
• Customer Communication & Education: Abandon accuracy measure and use more appropriate Precision, Recall & ROC curve to measure model performance
• Result: Recruitment work-flow enhancement: Customer is able to make informed decision

• Building model to prepare dataset for Triplet search algorithm
• Built python generator to feed image samples to Neural Network for training on limited memory environment
• Used Google’s Inception model & transfer learning method

• Built LSTM on characters (for language modeling) &
• Seq-2-Seq LSTM model for Language translation