1. Project Title: Invoice Processing- Identified and processed invoices of three distinct types, each with internal variations, utilizing the LayoutLMv3 multimodal model.- Achieved an impressive F1 score of 0.97 through accurate detection of relevant fields within the documents.- Implemented resulting solution to reduce manual labor by 75 hours daily, significantly enhancing operational efficiency.- Deployed the solution using cloud functions for seamless integration and scalability.2. Project Title: Auto Reply Suggestion- Analyzed seller and customer chat data to generate standardized reply suggestions.- Developed standard seller responses based on common interactions for broad applicability.- Employed a zero-shot classification model, mDeBERTa, to detect customer message intent.- Utilized a mapping system to link intents with candidate responses.- Calculated cosine similarity between candidate responses and customer messages using a clips/mfaq model to provide the top 3 responses with the highest score.- Achieved a mean similarity score of 0.98 between suggestions and actual responses, with a manual validation accuracy of 0.94 for relevance.3. Project Title: Master SKU- Conducted catalog cleaning by grouping data based on regex-cleaned titles and SQL grouping.- Utilized STSB model to compute embeddings of SKU titles and performed clustering to group similar SKUs.- Created review pairs for clustered titles, manually labeling them as identical or different.- Fine-tuned STSB model using contrastive loss to generate embeddings tailored to project data.- Successfully consolidated 53M SKUs into 36M master SKUs using the outlined approach.

- Developed a digitization module to do OCR on car RC card for the validation process of used cars which decreased the manual inspection efforts by 79%- Led a POC to create a car inspection product, successfully built models for car orientation classification, car and car part segmentation, and damage detection using CNN architectures like MaskRCNN, UNET, DeeplabV3, and also SVM trained on top of HOG features. Achieved an IoU of 0.83 on segmentation task- Initiated the development of a standard annotation tool for the organization and later deployed it on a server- Contributed to creating APIs and deploying their DOCKER containers on the server to serve rps of 165- Co-ordinated with cross-functional teams to make sure smooth incorporation of Data Science solutions

- Co-ordinated with cross-functional teams to carry out development activities for web pages using Angular 7- Developed a POC for a new client by communicating with senior members and presenting the solution to them- Upgraded version of angular in the application from version 7 to 9, making sure all existing features are working perfectly. This reduced the size of the application by 9% and thus improved performance- Made the application compatible with legacy browsers like IE11- Involved in accelerating lagging unit test cases of the project to achieve a code coverage of more than 80%