Projects

A collection of experiments, tools, and applications I've built exploring the intersection of web technologies, systems, and AI.

FAQ Chatbot with Embeddable Widget

Production-ready AI-powered FAQ system with embeddable widget for any website. Implements RAG architecture using LangChain, Groq LLMs, and Google Gemini embeddings with Chroma vector store for semantic search. Features real-time streaming responses, multi-language support, and customizable theming through data attributes. Enterprise-level architecture with iframe-based embedding via single script tag.

FastAPI LangChain Groq API Gemini Chroma JavaScript RAG Oracle Cloud

Personal Portfolio & Interactive Resume

Modern, responsive portfolio website featuring interactive particle-based portrait system with RGB chromatic aberration effects. Built with vanilla JavaScript and HTML5 Canvas for optimal performance. Includes AI-powered chatbot integration, smooth animations, and a clean, minimalist design optimized for 60fps rendering across all devices.

JavaScript HTML5 Canvas CSS3 Responsive Design AI Chatbot

Autonomous Driving: Lane Segmentation & Traffic Sign Detection

End-to-end perception system for autonomous vehicles combining semantic lane segmentation and multi-class traffic sign detection. Trained YOLOv11x-seg model on custom-labeled dataset with 3 pixel-level classes and 16 traffic sign categories. Achieved real-time inference capability with robust performance across varied lighting and road conditions.

Python YOLOv11x-seg Ultralytics OpenCV Computer Vision

Automatic Short Answer Grading Using NLP

NLP-based system to automatically grade open-ended exam responses using contextual embeddings from BERT. Fine-tuned pre-trained transformer models and built complete pipeline for text preprocessing, tokenization, and scoring. Applied ensemble methods to improve robustness across different question types.

Python BERT Hugging Face PyTorch NLP

Turkish-English Machine Translation with UMT5

Bidirectional translation system using Google's multilingual UMT5-small model. Fine-tuned on TED2020 parallel corpus (~80k training pairs) achieving 12.57 BLEU score. Built user-friendly Streamlit interface and monitored training with Weights & Biases. Implemented dynamic prefixing and tokenization pipelines for robust translation.

Python UMT5 Transformers Streamlit NLP

More Projects Coming Soon

I'm constantly experimenting with new technologies and ideas. Follow my work on GitHub to stay updated.