Prabhat Kumar

Risk Control Chatbot and Data quality:

• Built Question Answer and classification model using BERT and Roberta architecture.
• Fine-tuning of dense passage retriever and reader model using haystack framework and without haystack.
• Distillation of model and creation of student model using model architecture like TinyBert.
• Experienced in building models like intent and entity detection models for risk domain.
• Experimented with complex architecture like siamese network for text and intent similarity.
• Experimented and fine-tuned RASA's DIET Classifier.
• Well versed with text augmentation techniques based on word embedding, contextual word embedding, sentence embeddings etc.
• Experimentation with head of model creating different experiments based on pooled and sequence output of model
• Took leverage of multi-GPU environment of AWS and experimented with different distributed training strategies like Dataparallel, and DistributedDataparallel.
• Worked extensively on Jenkins, Kubernetes and is well-versed with CI/CD.

• Worked on projects namely: Project UNO, iVOC, Intelligent Banking, and Cheque Analytics
• Built pyspark-based recommendation engine for recommending personalized offers to customers with intention to cross-sell products, increasing existing product usage and converting dormant customers to active ones
• Built models like Next Best interaction (Acquisition model, Usage model, and retention model), time model, channel model, and content model as part of recommendation engine
• Built risk score models and worked on constrained optimization problems (operation research). Also performed analysis of problem for identification of objective function and associated constraints
• Experienced in operation research use cases and have working knowledge of PuLP, pyomo, and google or.
• Created pyspark-based python package for data creation, feature selection, parallel hyper tuning, parallel training, evaluation, and inference
• Experimented in modelling with xgboost, catboost, mmlspark lightgbm, lightgbm (non-spark), and neural net architectures like tabnet, tab transformer, node, etc
• Built image captioning model for content model and conducted experimentation with two-stage (based on faster R-CNN and transformer) and single-stage (fully transformer-based model) architecture
• Involved in A/B testing of models for recommendation engine/predictive models
• Worked on NLP uses case for sentiment analysis, review categorization, and article recommendation.
• Extensive knowledge of transformer architecture like bert, Roberta and experience in experimenting with different layers of architecture. Also experimented with creating different heads using pooled and sequence output
• Created experiments for doing unsupervised learning and then reusing layers in supervised learning with labeled data
• Experimented with various text augmentation techniques and check survival for models by conducting adversarial attacks
• Utilized NLP techniques like zero-shot learning in content model and feature discovery phase.
• Experienced in using hyper-parameter libraries like optuna, hyperopt in both non-spark and spark environments for traditional ML and deep learning use cases
• Experienced in end-to-end use case formulation to development of model till production deployment.
• Worked extensively on Jenkins and is well-versed with CI/CD.
• Involved in people management tasks like mentoring and guiding resources, taking feedback, and ensuring that their career is on right path.
• Ensuring team members are not facing any showstoppers and are on right track to complete assigned task.
• Also involved in various POCs related to NLP, Classification, and Speech Analytics and also participated in various hackathons on Analytics Vidhya and HackerEarth ranging from topics like classification, regression, image classification, and NLP
• Created revenue lift of 25 million SGD with Next Best Interaction model.

• Worked extensively on ETL projects where task is to automate various process using python
• Worked extensively on performance engineering and suggested fine-tuning of SQL scripts
• Creating C scripts and deploying them to HP Performance center for performing load testing on servers
• Use statistical methods for extrapolating performance metrics of Tech test server to production servers
• Experienced in developing PL/SQL code for generating data for various reports.

• Worked on US Healthcare domain
• Analysis of error on production and perform replication in QA environment
• Analyze new requirements and provide analysis documents to developers for coding
• Creating Business Acceptance test-case for final acceptance
• Involved in reviewing code and suggesting any changes if required.

• Extensively worked on AS/400 (iSeries system), PL/SQL, SQL, and campaign management in telecom domain
• Implemented KPIs using SQLs and PLSQL code
• Implement KPIs to build segments and campaigns
• Tracking performance of campaigns and optimizing them if needed