My work mainly involves developing deep learning models for extracting relevant information from documents and deploying them to production for inference. Here are a few highlights -- Worked on re-architecting the classification pipeline in AWS to parallelize the processing of pages in a document. The pipeline consists of 8 ML models working in tandem to classify and sub-classify the pages. It has helped our system to scale better with new customers onboarding.- Fine-tuned and deployed several LayoutLM (BERT-like) models for information extraction from different document types. The process involves data collection, cleaning (if required), pre-processing, training and spinning up AWS resources for deployment.- Trained a common Yolov5 model for detecting signatures from all documents. Applied techniques like synthetic data generation and two-strike training to handle edge cases.- Continuous collaboration with QA, Product, Data and Software teams to ensure all the stakeholders are involved and we produce the highest quality features. Also, involved in the hiring of new ML engineers

Worked with Prof. Vasileios Tzoumas on robust outlier detection for computer vision applications. We are developing a methodology to efficiently eliminate outliers from input sensor data without any prior knowledge of noise distribution or threshold.

I worked with the Enterprise Networking team and was responsible for development of software stack for their routers/switches. I was involved in following activities : - Design of software architecture and implementation of features currently being used in customer's networks. - Writing and executing exhaustive test cases before shipping the features to customers. - Debugging customer issues.

Worked on Spiking Restricted Boltzmann Machine trained using event based contrastive divergence based on Spike Timing Dependent Plasticity rule for weight update.

Built a Robot Operating System (ROS) framework for autonomous landing of Unmanned Aeriel Vehicle (UAV) on the target location using computer vision techniques.