Nov 2020 – Apr 2023 · 2 yrs 5 mos · Greater Bengaluru Area

• Using ML, AI at scale to find needle in haystack issues in the delivery network for Prime Videos. Working on streaming data of the order of 20million points per min.

Jul 2020 – Oct 2020 · 3 mos · India

• Working on AI problems related to CyberSecurity space
• Technologies : Python, Trax, Tensorflow, NER, LSTMs, AutoEncoder, Siamese Network Architectures, AutoML, Pycaret, Elastic Search, Flask
• 1. CyberIncident Allocation
• To automatically assign user to an incident based on historical data for incident and users to optimise incident resolution
• Productionized it as an ML service by developing an AutoML pipeline.
• 2. Related Incidents to an Incident
• a) Implemented an AutoEncoder to reduce the dimensions of cyber incident features like IOCs, Malware Family, Thread Actors, Tactics, Techniques, BUs, location etc.., to a latent space.
• b) Implemented a modified Siamese Network Architecture to compute vectors for text and description of Incidents and the latent vector from AutoEncoder to generate a model to find Incident - Incident similarity
• c) Exposed this as as ML service that gets the top 5 similar incidents to an incident integrated with the front end of the product
• 3. Connect The Dots
• Objective : To find related STIX entities to a cyber security incident
• 1) Implemented an NER tagger using LSTM architecture to tag entities like Malware Families, Threat Actors, Tactics and Techniques in Articles and Incident Text related to cyber security
• 2) Then Created a Knowledge Graph to generate connection between the entities using Elastic Search
• 3) Exposed this as an ML service to generate connected entities to an incident

Jul 2019 – Jun 2020 · 11 mos · Bengaluru Area, India

• Collection Agent to Customer Allocation
• Collection Agents are people who call customers when their loan goes into arrears. Objective of the project was to build an AI to intelligently allocate agents to customers, so as to optimise on the resolution rate of loans in arrears.
• We started with building a data pipeline using Pandas and Mysql to extract more than 100 indicators for a customer, loan and agent from multiple data sources.
• We then trained an XGBoost model after preprocessing the above data. Hyperparameter tuning was then performed using Amazon Sagemaker, to identify the best parameters of model, evaluating it on a validation set.
• We obtained a train and validation AUC of 73% and a test AUC of 72%.
• We then built a pipeline for model inferencing and analysed the scores produced by model on various agents mapping it back to the loans parameters to see the models performance from a business point of view.
• Overall business impact of the project was estimated to increase collections revenue by 10%.
• Finally, we built an allocation strategy to account for the workload distribution of the agents and the scores produced by the model across agents.
• Bank Statement Classifier
• Objective - To build an AI to automatically validate Bank Statements submitted by customers during KYC process.
• A classifier was built using Keras-Tensorflow and trained on 30000 images of Bank Statements and 100000 images of Non Bank Statements (comprising of other document types - Forms, Invoices, Emails etc.. from RVL-CDIP dataset)
• After multiple experimentations with dataset, models, hyperparameters... we increased the benchmark performance of 88% AUC to 96% AUC and 97% Accuracy on an out of sample test set.
• A significant boost was obtained by introducing oversampling of positive samples and error analysis of the misclassified points.
• To speed up training, we trained the model on a GPU using Cuda libraries.

Jul 2018 – Jun 2019 · 11 mos · Bengaluru Area, India

• Solved problems related to recommendation of product categories to users in retail space.
• Worked on user segmentation problem based on buying behaviour of users.
• Worked on the problem of Message mining i.e. Prediction of 3 day hit rate of message templates for SmS Campaigns
• Technologies : Spark, Scala, Python, Recommendation Systems, Doc2Vec, Paragraph2Vec, Topic Modelling, Text Mining

Dec 2016 – Jun 2018 · 1 yr 6 mos

Jun 2014 – Dec 2016 · 2 yrs 6 mos

Jun 2013 – Sep 2013 · 3 mos

• Analyzed clock data from different servers to estimate offset and frequency correction required to correct the system clock upon start-up.

Jan 2013 – Jan 2013 · 0 mo

• I worked on a project on Online Character Recognition. I designed some heuristics to pre-process the data to reduce the errors. Then designed a feature set to train and test the model (kNN) on the pre-processed data. With similar user test sets obtained accuracy of 92%, with a different user test sets obtained an accuracy of around 75%.

Apr 2012 – Jul 2012 · 3 mos

• The project dealt with application of Natural Language techniques on Web Query Language for Augmenting the Query logs and Improving the overall search efficiency of Search Engines.

Jul 2009 – Jul 2014 · 5 yrs · Kharagpur I, India

• Completed my Masters and Bachelors in Computer Science and Engineering at IIT Kharagpur.