Docupilot

Docupilot is an advanced AI-driven web application built for the Round 3 Experience and Insight Prototype of the Adobe India Hackathon. It transforms static PDF reading into an interactive, insight-driven journey by combining document understanding with deep analytics, semantic navigation via the Adobe Embed API, and automated conversational AI podcast generation.

A Review of Advances in Large Language and Vision Models for Robotic Manipulation: Techniques, Integrations, and Challenges

Recent advancements in transformer-based systems, including Large Language Models and Large Vision Models, have significantly transformed robotic manipulation by enabling enhanced task planning, real-time decision-making, and adaptive behaviour in complex environments. This review synthesises current research on integrating these models with robotic control systems, highlighting innovative strategies that merge linguistic and visual processing to improve precision and efficiency.

Linguistic Decoding RL

Linguistic Decoding RL is a reinforcement learning environment that trains LLM agents to perform long-horizon hidden state inference—decoding latent intent and true risk from biased text and temporal behavior. Built around high-stakes Indian MSME and startup contexts, the platform uses GRPO to teach a Qwen 1.5B model to cut through surface-level statements, track financial and behavioral drift across multi-step episodes, and execute appropriate interventions.

The Brittleness of AI-Generated Image Watermarking Techniques: Examining Their Robust-ness Against Visual Paraphrasing Attacks

The rapid advancement of text-to-image generation systems, exemplified by models like Stable Diffusion, Midjourney, Imagen, and DALL-E, has heightened concerns about their potential misuse. In response, companies like Meta and Google have intensified their efforts to implement watermarking techniques on AI-generated images to curb the circulation of potentially misleading visuals. However, in this paper, we argue that current image watermarking methods are fragile and susceptible to being circumvented through visual paraphrase attacks.

LiveCold

LiveCold is an advanced, real-time cold chain intelligence platform built on the Pathway streaming framework, designed to tackle India's ₹92,000 Crore annual losses from perishable supply chain failures. Moving beyond passive monitoring, LiveCold processes high-velocity IoT telemetry in milliseconds to autonomously detect anomalies, evaluate financial trade-offs, and execute cargo diversions before spoilage occurs.

ParkSight

ParkSight is a lightweight, CPU-only, pre-emptive parking enforcement dispatcher designed for the Flipkart Gridlock 2.0 Hackathon. It transforms reactive parking patrols into a predictive, data-driven operation by analyzing historical violation records to detect statistical hotspots, forecast high-risk zones, and dynamically route enforcement units using OR-Tools and a Streamlit command-and-control dashboard.

The Visual Counter Turing Test (VCT2): A Benchmark for Evaluating AI-Generated Image Detection and the Visual AI Index (VAI)

The rapid progress and widespread availability of text-to-image (T2I) generative models have heightened concerns about the misuse of AI-generated visuals, particularly in the context of misinformation campaigns. Existing AI-generated image detection (AGID) methods often overfit to known generators and falter on outputs from newer or unseen models. We introduce the Visual Counter Turing Test (VCT2), a comprehensive benchmark of 166,000 images, comprising both real and synthetic prompt-image pairs produced by six state-of-the-art T2I systems.

Flood Segmentation

Flood segmentation is a deep learning model that segments flooded area from aerial images.

CalmSpace

CalmSpace is an end to end sentiment analysis platform where the user can record or upload the recorded files of their emotion and the deep learning model will learn through algorithm like RNN Model Architecture, and will predict the emotion of the voice . The sentiment analysis will be shown through various graph plot in much understandable way.

DiagnoAI

DiagnoAI is a tool to detect a disease from a text description of the patient's symptoms and daily condition. It is based on a transformer model called BERT, fine-tuned for 24 common diseases.

Clean/Dirty Road Classifier

Project is based on ConvNets and Transfer Learning, and is used to identify whether a road is clean or dirty. MobileNet was used as our base architecture and the weights are based on imagenNet. Web Scraping techniques were used to download images of both clean and dirty roads from the internet. The results of this model was pretty good, with a training accuracy of 97.1% and validation accuracy of 91.6%.

RoadMent

An end-to-end project based on Image Segmentation. A website was developed, where if a useruploads a satellite image the model will segment out the roads from the images. The U-Net architecture was used for our model and it was a group project. Main role here was to handle the data, build the model and fine tune it accordingly. Model perfomance was based on IoU score and the loss was Dice-Coefficient loss. Training IoUscore was 0.497 and validation IoU score was 0.415. It was the 2nd Best Project from East Zone in Anveshan National Student Research Convention 2021-2022 and was in Anveshan representing NIT Silchar.

Autonomous driving application Car detection

This project was a Keras model that detects objects such as cars, buses, pedestrians, etc which are in the image taken by the front camera of the car. This project is made using YOLO algorithm and Convolutional Neural Network in deeplearning.ai specialization on Coursera.

Emotion Detection in Images of Faces

This was a Keras model that does a happy/sad classification of images of faces. This project was made using Convolutional Neural Network under deeplearning.ai specialization on Coursera.

Textual Entailment as an Evaluation Metric for Abstractive Text Summarization

This project is an NLP model which summarizes text using technique Abstract Summarization with the help of Textual Entailment.

Segmented sequence modeling in Indian classical music

This paper is published under IEEE Silicon Conference 2022. The idea of learning the domain knowledge using Variational AutoEncoder(VAE) to detect the stated landmark in Indian Classical Music(ICM).

Grad-CAM for the skin-mnist dataset for skin lesion diagnosis

This project uses a convNet model for detection of skin lesion detection and uses Grad-CAM for explaining the prediction of the model. It uses a modified version of the skin-mnist dataset which only contains the classes nv, bkl, mel (around 550 images of each).

Cough it

This project takes a deep learning approach to analyze the acoustic dataset containing cough sound recordings of both COVID-19 positive and negative individuals. It proposes a ConvNet model that performs the classification between COVID-19 positive and negative with a notable AUC score of 87.07 on the blind test set provided by the same for unbiased evaluations of the models. The model takes in 15 MFCC features of the sound examples as input and produces the probability score of the classification as output.