Kevin Godfrey Verpula

Combined VGG16 convolutional feature extraction with Vision Transformer attention mechanisms for CIFAR-10 classification. Achieved 4% accuracy improvement by locking VGG16 layers and feeding outputs to Transformer's non-encoder layers.

Developed steganalysis system using Inception V1 (GoogleNet) to detect hidden messages in images. Improved detection accuracy by 6% with ensemble learning and gradient boosting, achieving 85% overall accuracy.

Developed CNN for complex datasets including CIFAR-10 and binary classification (cat/dog recognition). Integrated VGG16 for performance comparison, achieving 3% accuracy improvement over traditional CNN architectures.

Built decision tree classifier handling numerical and categorical data, applied to UCI Weather dataset achieving 95% accuracy. Implemented both Gini impurity and entropy-based splitting criteria for optimization.

Designed RNN and LSTM model for sentence completion task by predicting the most probable next word in a sequence. Demonstrates practical application of sequence modeling and probability-based text generation.

Implemented GAN to generate realistic synthetic images using MNIST dataset. Model progressively improved image quality through adversarial training, achieving remarkable resemblance to original dataset.

Utilized BERT for extractive question-answering system. Successfully extracted direct answers from passages but showed limitations with indirect or implied questions, highlighting areas for fine-tuning improvement.

Implemented fully functional neural network from ground up with forward propagation, backpropagation, and gradient descent optimization. Built to gain comprehensive understanding of neural network architecture fundamentals.

Created VR-based games and simulations for construction worker safety training. Replicated daily construction scenarios with automated gear handling and integrated haptic feedback for immersive, realistic training experiences.

Developed educational game showcasing mathematical concepts (linear algebra, geometry, calculus) in interactive format. Incorporated reinforcement learning and A* pathfinding for real-time educational gameplay.

Built hybrid React-Node.js interface to streamline automated testing for remote devices. Integrated Jenkins for pipeline management and Selenium for end-to-end testing, ensuring efficient and reliable automation.

Developed web application for Karunya Hacks hackathon to reduce food wastage by connecting restaurants with those in need. Platform allows food providers to list surplus food for low-cost delivery, promoting social responsibility.

Developed core UI and operating environment for RDK STBs featuring Power Management, USB functionality, custom media player, TV schedule menu, XCast casting capabilities, and integrated Alexa & Google Assistant voice controls with third-party app compliance.