VIVEK MANOHARAN

DocumentAI Workbench [Python, GCP, DocumentAI, Flask, Apache Beam, Protobuf] - Developed parsers for identity documents, lending and tax forms using DocumentAI’s custom document extractor. - Added feature improvements to fake document synthesis pipeline and template creation application - flask based application hosted on Google App Engine which creates templates for creating synthetic documents. - Ran experiments to demonstrate the effect of models trained using synthetic documents vs real labeled documents, achieved a F1 score of 0.86 using noise added synthetic documents.

Document Extraction Solution [Python, Tensorflow, CNN, Faster R-CNN, SSD, SpaCy, OpenCV, Tesseract OCR, Flask, Docker] - Led design and development of entity extraction solution for tabular structured data using objection detection, image processing, NLP and template based approaches. The solution extracted entities from 70+ document variations with an accuracy of 0.85 which reduced the manual effort by 35%. - Developed multiple ner models for extracting non-tabular structured data from 20+ documents. The models were exposed as REST APIs using TF Serving. - Developed CNN based image classifier to identify different formats of documents, cheques and summary pages. - Wrote image pre-processing scripts to clean documents and improve data capture with OCR. - Set up an on-prem application pipeline by containerizing the modules and exposing them as REST APIs. - Developed a labeling tool for ner models leveraging Tesseract OCR’s hOCR values.

Ticketing and User Management Applications [C#, ASP.NET MVC, SQL Server] - Migrated application features from windows-based application to web application. - Involved in end-to-end SDLC, integrated new application features and implemented bug fixes. - Involved in production support and incident management, L3 support.