Senior Data Scientist and AI Educator with 4.5 years of industry experience, 2 years of academic teaching, and 5 years of applied research in advanced analytics, predictive modeling, deep learning, computer vision, NLP, and generative AI. Passionate about translating complex AI concepts into business-relevant learning paths for technical and non-technical audiences.
Feb 2023 — Present, Assamese OCR and sentiment analysis from degraded text documents, Starting from a collection of old Assamese script from religious and folk documents, this work involves in designing a OCR system to detect and binarize degraded texts., It is a two stage approach to detect the degraded regions in the document using CNN based approach and predict the next word or character for those parts using GAN and BLSTM based model., Technology: CNN, U-net, GAN, BLSTM, Pytesseract,
Mar 2023 — Feb 2024, AI-powered bot using NLP and computer vision models for automated report evaluation, The chatbot is designed to evaluate students thesis document based on certain parameters, including inclusion of required chapters, minimum requirement of explanation, number of articles referred, image quality, summarization of major chapters and checking results and discussion., Utilized Hugging Face’s Transformers library to load pre-trained NLP model GPT-3 and guided the model with specific prompts., Technology: Transformers, GPT, NLP, LLM, prompt engineering,
Jun 2022 — Jun 2023, developed an academic project for medical image segmentation and caption generation model using pre-trained transformer models, trained the transformer-based model for medical image segmentation, where the model learns to delineate and classify different structures or abnormalities within the medical images, fine-tuned the pre-trained transformer model using the medical image dataset to improve segmentation accuracy and generalization to unseen data, extended the transformer-based model to generate descriptive captions for segmented regions within the medical images, trained the model to generate informative and contextually relevant captions that describe the detected structures or abnormalities in the images
Jan 2018 — Feb 2021, 3D mesh surface models reconstruction using deep neural networks, Designed fragile mesh watermarking algorithm for 3D surface mesh models considering verification of a wider set of mesh geometrical as well as topological attacks., ML concepts were used during mesh segmentation and mesh reconstruction. For reconstruction of the models graph-CNN architecture is used., Technology: CNN, Transfer learning, Auto diffusion function, Octree data structures, Quantization index modulation, Prediction error Expansion.