Real-time Sign Language Action Detection System

• Developed a real-time Sign Language Action Detection System achieving 90% accuracy.
• Integrated face, hand, and body landmark detection for precise sign recognition.
• Trained LSTM neural network reaching 85% accuracy, improving performance by 25%.
• Enhanced accessibility for individuals with hearing and speech impairments.
• Enabled applications in education and assistive communication tools.

• Engineered a SOTA multi-document summarization model for news aggregation, achieving an 87% human acceptance rate with less than 1% critical errors.
• Designed automated news timelines for evolving events, resulting in a 2.8% increase in content depth and a 1.4% boost in user time spent.
• Optimized ranking algorithms with a CTR prediction model, increasing daily active users by 4.3% and user engagement by 2.1%.
• Created a novel similarity scoring formula, improving F1-score from 0.91 to 0.95 for news clustering.

• Used StarGAN architecture to classify images across multiple domains.
• Achieved an F1 Score of 0.83, Precision of 0.9917, and Recall of 0.8051 on the CelebA dataset.
• Trained model over 500,000 iterations using PyTorch and Adam optimizer.
• Applied adversarial loss, classification loss, and gradient penalty for balanced training.
• Optimized performance with specific hyperparameters ensuring stable performance.

TinderGPT is designed as an AI assistant to streamline the dating experience, sparing users from the time-consuming and often uncertain process of messaging on dating apps. This AI handles everything on behalf of the user—from initiating conversations with messages tailored to the profile in question, crafting engaging interactions, to arranging dates and providing the user with the match’s contact information.

In this project, I focused on implementing small open-source models, such as Zephir-7b, on single-board computers like the Raspberry Pi. This setup enabled the creation of a robot controlled by an LLM. Working with such compact models presents numerous challenges; one significant issue is the reduced level of intelligence compared to larger models like GPT-4. To address this, the project required fine-tuning the open-source LLM to achieve satisfactory results.