MrTimmyJ

Project Overview

        Dracula AI Agent is an interactive, full-stack AI chatbot trained on Bram Stoker’s Dracula. Originally built as a Hugging Face model, it has since been expanded into a production-ready FastAPI web service, hosted on an Ubuntu server with Nginx.

        The system combines a PyTorch-based GPT-style model (trained from scratch with tiktoken tokenization) and a frontend chat UI (“VampChat”) that allows users to converse with Dracula in real time. This project demonstrates expertise in machine learning, backend deployment, and AI-driven web applications.

🧠 Model Architecture

• GPT-style transformer with token and positional embeddings
• Trained with cross-entropy loss using AdamW
• Inference supports temperature/top-k sampling for varied outputs

📊 Tokenization & Data Pipeline

• tiktoken (GPT-2) encoding for efficient tokenization
• Text chunked into fixed-length context windows with stride
• PyTorch DataLoader batches/shuffles train/validation splits

🧪 Training & Evaluation

• Multi-epoch training with periodic validation passes
• Loss tracking and visualization for diagnostics
• Checkpoint save/load via torch.save / torch.load

🖥️ Backend API

• FastAPI endpoints (e.g., /generate) serve model inference
• JSON request/response schema for prompts and outputs
• CORS enabled for browser clients behind Nginx

💬 Frontend

• HTML/CSS/JavaScript chat UI (“VampChat”)
• Sends user prompts to the API; renders model replies (optionally streamed)
• Simple session controls (new chat/clear history)

🌐 Hosting & Deployment

• Deployed on Ubuntu with Nginx reverse proxying to the FastAPI app
• Nginx serves static frontend and proxies API requests
• Basic logging/monitoring for reliable production use

Project Features

• 📚 Custom GPT Model: Transformer-style architecture trained on full Dracula text.
• 🔡 tiktoken Tokenization: GPT-2 compatible encoding for efficient batching.
• 🔁 Training Pipeline: PyTorch training loop with checkpointing and evaluation.
• 📈 Diagnostics: Real-time loss visualization with Matplotlib.
• 🧛 Interactive Chat UI: Web frontend where users “talk with Dracula.”
• 🌐 Web Deployment: FastAPI backend served directly, with NGINX handling static frontend files.
• ⚡ Chat Integration: Frontend sends POST requests to FastAPI to generate Dracula’s responses in real time.

Project User Workflow

• Place the cleaned Dracula text into your data directory
• Run training with Python gpt_train.py
• Monitor training loss and visual plots
• Generate text samples using a provided prompt
• Save and load model checkpoints for further experimentation