• Integrated RAG for multiple languages and file formats(PDFs, CSVs, and excel). It's used for question answering from documents, analyzing data, and creating user-specific plots. Here we also extract tables and images to provide more information. Also working around large language models like GPT-4 and Claude 3 to add better capabilities. • Added new Inverse Prediction flow. This led to more accurate results for material properties prediction based on a set of ingredients.• Guided 2 interns throughout their time here and also helped in automating CI/CD and database workflow.

• Integrated Explainable AI for regression and classification models. Presenting multiple charts to understand predictions made by models. Deployed it on AWS Fargate for scalability.• Reduced response time for a Machine Learning API by 20%.• Worked on integrating classification models for imbalanced datasets and Data Sanity checks.

I taught my peers and juniors on how to use Tensorflow and Keras to build deep learning-based applications. Also guided them on building end-to-end projects using these frameworks.Hosted four challenges in building apps around Computer Vision and NLP.

• Developed a custom Object Detection + OCR model for reading Digital Odometer. The mAP for the detection model was 96%.• Worked on Instance Segmentation and Object Detection using Pytorch. Goal here was to detect the dents, scratches, scrapes, cracks in windshield etc. present on cars and estimate the repairing cost.• Integrated MLflow for model tracking and Data pipeline for Classification tasks. The classification model was trained for determining the different properties of the car. It included 30+ classes.• Worked on Detection Transformer(DeTR) for instance segmentation. Also worked on Video Analytics for getting useful insights from video.

• Mainly worked on Computer Vision projects using Tensorflow. It included Object Detection, background removal and knowing the color and brightness of an image.• Contributed to ongoing projects using YOLO for Object Detection and OpenCV for tracking and counting objects. This was used in manufacturing industries for counting bottles, towels and boxes. And mAP achieved here was 97.5%. During the Inference test video ran at around 50 FPS.• Got hands-on experience in annotating real-world industry data using LabelMe and LabelImg for the custom detection models.• Deployed 2 POCs using Streamlit and Flask. One project was to detect defects in fabrics and calculate the area covered by them. Another project was extracting tables from bills or receipts using OCR. Pytesseract was used for OCR.

Here I helped teams to build the ML projects for the hackathon. Projects were around Object detection, Pose Estimation, and predictive analysis.One of the teams I assisted made it to the finals. They created a project that helps people who can't speak by detecting signs and translating them into language.

• Worked on a framework named SpiceAI for building a predicting tweet likes based on the past history of the user.• SpiceAI used Reinforcement Learning to make predictions on given time-series data.

• First I learned about analyzing satellite images using OpenCV and Pillow.• Built a custom CNN model to detect lakes and rivers from satellite images and then used the company’s prebuilt system to identify its location.• Worked on setting up a pipeline for collecting and cleaning aerial & satellites data.