Projects

Neural RAG Engine

￮ Built a Retrieval-Augmented Generation (RAG) system with a custom PyTorch projection head, sentence-aware document chunking, and in-memory cosine vector search. ￮ Integrated Google Gemini for context-grounded answer generation with source citations, offline fallback, and an interactive web dashboard with REST API.

Music Downloader Bot

Telegram bot that accepts a song name, YouTube URL, or Spotify track link and delivers audio with metadata and cover art. Serving real users in production on ARM64 with zero-downtime uptime. ￮ Async PostgreSQL-backed task queue with concurrent workers — multiple users served simultaneously without blocking the event loop ￮ Two-phase download (metadata → size check → download) rejects oversized files before writing any bytes ￮ file_id caching layer eliminates redundant YouTube hits — repeat queries resolve instantly from DB ￮ YouTube datacenter IP blocking solved by routing through Embed-based search instead of direct extraction ￮ Spotify track resolution via Open Graph scraping — no API key, no extra dependencies ￮ Deployed with systemd auto-recovery, health-checked DB pooling, ephemeral storage cleanup, and specific user-facing error messages for every failure path

Image Classification: Dogs vs Cats

Developed a TensorFlow/Keras convolutional neural network for cat vs. dog image classification, including data preprocessing, model training, and evaluation, achieving 89.8% validation accuracy on the Kaggle Dogs vs. Cats dataset.

GoSQL Interface

A web-based admin panel built with Go and PostgreSQL for managing users, network interfaces, and ACL rules. Features authenticated access with role-based permissions, searchable tables, modal-based CRUD operations, and a schema viewer. Designed as a practical learning platform for database connectivity, authentication workflows, and full-stack development with Go, HTML templates, Vanilla JavaScript, and PostgreSQL.