Hi, I'm Keval

WorldQuant Brain – Alpha Research

Ongoing

Technologies: Alpha · Statistical Modelling · Factor Research

• Active participant researching and submitting quantitative alphas across global equity markets.
• Developing systematic trading signals using price/volume data, fundamental factors, and statistical techniques.
• Iterative backtesting and alpha performance refinement on the WorldQuant Brain platform.

Carbon Campus

Feb'24 - May'24

Technologies: REST API, Dart, Flutter, MongoDB

• Developed a mobile app to calculate carbon footprint based on user input like energy, food, vehicles and flights.
• Features a chatbot, powered by the Gemini API, offering personalized suggestions for reducing carbon emissions.
• Visualizes carbon emissions of college buildings, including mess halls, canteens, and other infrastructure, displaying real-time data.
• Includes a carbon trading system where users can trade based on their emission amounts.
• Displays real-time news updates related to carbon emissions and sustainability trends, with integration of the Wren page.

SecureVue AI

Dec'23-Feb'24

Technologies: LSTM, CNN, LRCN, OpenCV, ReactJS

• Developed as part of the GDSC Google Solution Challenge, the project focuses on advancing SDG 16.
• Engineered an AI model for Real-Time CCTV Surveillance to bolster efforts in enhancing security and monitoring.
• Designed an user-friendly video posting interface and a dynamic community forum for idea exchange.
• Incorporating the capabilities of CNNs and LSTMs, the system detects criminal activities with 80% accuracy.
• Applied frame prediction to classify crime types (arson, shootings, road accidents) with 99% accuracy.

🚨 Disaster Alert System

A responsive web application designed to enhance public safety by providing real-time alerts based on the user's current location. The system classifies alerts into three severity levels and dynamically adjusts based on disaster type and duration.

Technologies: JavaScript, JSX, Tailwind CSS

• Implemented a three-tier alert system (Red, Yellow, Green) for different severity levels
• Developed secure user authentication system with personalized alert preferences
• Created an Emergency Information page with safety guidelines for various natural disasters
• Built real-time location-based alert system with dynamic updates
• Designed responsive UI with Tailwind CSS for optimal user experience across devices

Guardians of Truth (Deepfake Detection and Fake News Check)

Dec'23-April'24

Technologies: CLIP-ViT-L/14, Resnet50, CNN, Fine-Tuning, LSTM, Javascript, FastAPI

• Implemented research paper using ProGAN dataset and CLIP (ViT-L/14) for feature extraction; achieved 92.9% accuracy on DALL·E and 87% on Glide-100, surpassing prior SOTA of 86% and 79%.
• Fine-tuned ResNet-50 for deepfake detection (87% accuracy) and deepfake source attribution (93% accuracy).
• Built TruthGuard for fake news detection, achieving 92% accuracy utilizing LSTM for sentiment analysis.
• Integrated Guardian for Deepfake and TruthGuard model into a web platform for real-time updates and verification.

🎮 AI-powered Stone-Paper-Scissors Game

This project is an interactive implementation of the classic Stone-Paper-Scissors game using computer vision and Python. The system detects the player's hand gesture in real time using OpenCV and CVZone, classifies it as Stone, Paper, or Scissors, and pits it against the computer's randomly selected move. The game then calculates the result and updates the score dynamically.

This project demonstrates the integration of real-time image processing, gesture recognition, and basic game logic, providing an engaging and educational example of how AI and CV can be used to bring traditional games to life.

Technologies Used: Python, OpenCV, CVZone, MediaPipe (optional), NumPy

📊 NAKAL: A Unified Training Framework for Table Recognition

A novel framework designed to streamline and unify the process of table recognition (TR) by combining a task-agnostic architecture, a unified training objective, and a consistent training paradigm.

Technologies: PyTorch, torchvision, tokenizers, PIL, BeautifulSoup, Matplotlib, IPython, HTML, Python

• Developed a unified framework that tackles three core TR tasks: table structure extraction, cell content recognition, and cell bounding box prediction
• Implemented self-supervised pretraining (SSP) on diverse unannotated tabular images
• Created a single language modeling objective to handle multiple TR tasks
• Designed a task-agnostic architecture that relies solely on pixel-level inputs
• Achieved scalability, robustness, and simplicity in extracting structured data from tables