Created a real time interactive application that recognizes both voice and phone movement patterns enabling interactive use cases such as music based games Developed backend services with FastAPI and Node js handling API communication and real time user interactions Integrated the Shazam API for music recognition ensuring responsive and scalable performance The tech stack included JavaScript Node js FastAPI and the Shazam API

Developed a CNN-based image classifier to identify over 120 dog breeds incorporating Grad-CAM for interpretability and visual explanations Trained on a dataset of 5000+ images and evaluated the model using TensorFlow Keras achieving 85% accuracy on validation data Enhanced usability by generating heatmaps that highlight the most relevant regions in the image for each prediction Tech stack included Python TensorFlow Keras and OpenCV

The System for Classifying Podcasts is an AI-powered Telegram bot that automatically categorizes podcast episodes by theme using Large Language Models (LLMs). The system integrates audio transcription with semantic analysis to provide users with summaries, recommendations, and insights such as main topics, target audience, and listening difficulty.

I designed and implemented the full pipeline , from audio retrieval via Spotify and YouTube APIs to transcription with Faster-Whisper and topic modeling using spaCy and HuggingFace models. By processing over 200 podcast episodes, the system achieved accurate classification into more than ten thematic categories, reducing manual labeling effort by 90%. The project utilized an open-source Llama Maverick 4 model via OpenRouter API, making it efficient and fully automated.