I build intelligent ML systems that ship.

ApollodB: Psychoacoustic EQ Generation System

• Built an open-source TensorFlow-based system on Google Cloud for psychoacoustic EQ generation, achieving state-of-the-art results on the DEAM dataset.
• Developed advanced neural networks to generate personalized audio equalization profiles based on psychoacoustic principles.
• Implemented cloud-based deployment using Google Cloud Platform for scalable real-time audio processing.
• Created a comprehensive demo interface for interactive testing and validation of EQ generation capabilities.
• Open-sourced the project to contribute to the audio engineering and machine learning communities.

Generative Networks and VAEs to Assist Artists

• Designed Variational Autoencoders (VAEs) for sketch refinement, improving workflow efficiency by 30% and enhancing creative control.
• Achieved 90.6% accuracy with the CUHK Face Sketch Database, presenting findings at ICCET 2023 in India.
• Implemented advanced techniques to retain finer details in generated images, significantly improving visual fidelity.
• Built a user-friendly interface for artists to seamlessly integrate the model into their creative workflows.
• Validated real-world usability with artists, receiving positive feedback on practicality and output quality.

Energy Price Forecasting Using LSTMs and ARIMA

• Developed LSTM and ARIMA models for energy price forecasting, improving prediction accuracy by 18% for Australian energy datasets.
• Conducted NLP-based trend analysis to identify hidden patterns, providing actionable insights for stakeholders.
• Improved model training speed through data preprocessing techniques, enabling real-time forecasting capabilities.
• Built visual dashboards using Tableau, simplifying interpretation for non-technical decision-makers.
• Presented findings at the AIJR Conference 2022, highlighting the balance between model performance and scalability.

• Created a personal LLM application fine-tuned on a Final Fantasy character dataset for daily assistance.
• Allows users to select their favorite character as their AI helper for a personalized experience.
• Utilizes the totl llama framework for efficient model serving and interaction.
• Supports context-aware responses and task automation tailored to user preferences.
• Designed for privacy, running locally without cloud dependencies.

• Automates web form filling using AI models and Selenium-based web automation tools.
• Generates context-aware answers for form fields using advanced language models.
• Designed for offline operation to maximize privacy and efficiency.
• Integrates Python 3.x, Chrome, ChromeDriver, and Ollama with LLaMA 2 7B for robust automation.
• Ensures compatibility and easy setup across different environments and browser versions.

• Utilized Pix2Pix GAN for transforming traditional anime images into enhanced, high-quality artworks.
• Trained the model on a diverse dataset of anime images, achieving superior color correction and detail preservation.
• Implemented an iterative feedback mechanism to refine output aesthetics based on user suggestions.
• Presented the project at a local tech fair, receiving commendations for blending technology and creativity.
• Optimized data preprocessing pipelines, reducing GPU training time by 20% without compromising output quality.

• Built an app to identify all 151 Gen 1 Pokémon and 15 mega evolutions using DenseNet201 and FastAI over PyTorch.
• Provided detailed information and stats for each identified Pokémon.
• Utilized advanced image classification for high accuracy and reliability.
• Designed a modern, user-friendly interface for quick Pokémon lookup.
• Optimized for mobile and desktop use, ensuring accessibility for all users.

• Developed a web application based on Wasserstein GAN with Gradient Penalty (WGAN-GP) to generate realistic paintings.
• Trained the model on a dataset of 8,000 Albrecht Dürer paintings for high-fidelity artistic output.
• Implemented a super-resolution mode to enhance generated images for detailed viewing.
• Designed an intuitive user interface for generating and downloading artworks.
• Optimized model performance for real-time generation and minimal latency in the web app.

• Developed a real-time hand gesture recognition system using a standard computer webcam.
• Implemented convolutional neural networks (CNNs) for accurate gesture classification.
• Enabled sign language recognition and automation tasks based on detected gestures.
• Optimized the model for low-latency inference and robust performance in varied lighting.
• Designed a user-friendly interface for live gesture feedback and system control.

• Developed a firearm detection application using OpenCV and YOLOv3, achieving high precision and recall metrics.
• Trained on a custom dataset of images and videos, optimizing the model for real-time security applications.
• Integrated the model with CCTV feeds for real-time detection, enabling proactive security monitoring.
• Optimized inference time through quantization and efficient model pruning, achieving a 25% reduction in latency.
• Tested the model under varying lighting conditions and environmental noise to ensure robust performance.