Souvik Majumder

Experience

• 

  Tata Consultancy Services

  Feb 2022 - May 2022

  Project Intern (Digital Cadre)

  📍 Worked in a team to develop an end-to-end project, "Video-Bin: AI-Driven Video Search" to solve a problem leveraging Artificial Intelligence, aiming at value creation for businesses in the media industry. • Implemented Natural Language Processing (NLP) algorithms like RAKE, YAKE, KeyBERT at the core of building the solution with Python as the primary programming language. • Designed the solution to be deployed on cloud, utilizing AWS S3, SQS, Lambda, Transcribe, DynamoDB services and Boto3 library. • Also developed a web application for users to interact with the solution. Show less

• 

  Siemens

  Jul 2022 - Nov 2023

  Machine Learning Engineer

  📍 Project for a Middle East client: Computer Vision based use case. • Processed the satellite image data and the Digital Elevation Map (DEM) layer data captured using LiDAR as part of the Data Engineering phase using the software QGIS. • Developed a custom Semantic Segmentation model. • Experimented and evaluated the performance of different model architectures like Unet, LinkNet, FPN, etc. combined with pre-trained encoders like resnet, mobilenet, efficientnet, etc. to build a custom model scoring 98.5% on the performance metric/business KPI. • Developed the automated Training and Inference Pipelines along with the corresponding Unit Test scripts to be integrated with SonarQube for continuous code quality analysis. • Researched and developed a new customized Monocular Depth Estimation modelling approach to generate 3D depth maps from single 2D satellite images to estimate real-world terrain elevation. Filed a patent for the same.📍 Project "Electric Load Forecasting": Research work on developing a state-of-the-art Deep Learning based Time Series Forecasting model for short, medium, and long-term load forecasting.📍 Project "Automotive Defect Analysis": Computer Vision use case in Assembly Line Automation for a vehicle manufacturer. • Trained an Object Detection model like Faster R-CNN to detect automotive parts located in the engine area of cars. • Trained and optimized a custom Image Classification model on the detected automotive parts to separate the perfectly aligned ones from the misaligned. Show less

• 

  Zensar Technologies

  Nov 2023 - now

  Senior Machine Learning Engineer

  📍Retrieval Augmented Generation (RAG) with complex data (HTML & PDF) for South African government • Devised a LLM powered RAG solution for users to chat with a bot based on information in the form of complex data (html or pdf) that contain tables and graphs embedded in them. • Fine-tuned and integrated NeMo Guardrails to restrict unwanted out-of-topic conversations, jailbreaking attempts, use of abusive language, and hallucinations. • Leveraged other Generative AI tools like Langchain, LlamaIndex, Unstructured IO, and Chroma DB to develop the solution. • A complete Azure based version of the solution has also been built to deploy for clients, replacing some of the open-source frameworks with their Azure counterparts like Azure AI Search as the vector db and Azure Prompt Shield instead of NeMo Guardrails.📍Insight AI Diagnostics (Medical) • Leveraged and integrated both Generative AI and Deep Learning Computer Vision models using LLM Agents to build a platform that enables - Classifying diseases analyzing medical images like X-rays, MRI and CT scans. - Performing segmentation to calculate disease spread and severity. - The users to communicate in natural language with the underlying fine-tuned AI models responsible for the diagnosis. • Two examples of deployed solutions are Object Detection and classification to detect Wilson's disease from brain MRI Scans, Semantic Segmentation to detect and localize polyp from colonial tract images.📍Image and Video Advertisement Generation Platform for a South African client • Leveraged open-source text-to-image and image-to-image generation models like Stable Diffusion to enable users in generating their own creative advertisements with images, videos and texts from both textual and image prompts. Show less