🧠 Building a Real-Time AI-Powered Mario Dashboard (With Achievements, Stats & More)

Imagine watching Mario play himself — learning from experience, tracking performance, and celebrating every mushroom and level cleared like it’s a milestone. Welcome to my latest project: an AI-powered real-time dashboard that turns classic Super Mario Bros. into a living lab of machine learning, memory, and game analytics.

🔗 View the full project on GitHub

🎮 What Is It?

This project uses Python, RAM state parsing, AI logic, and a fully interactive Flask + Socket.IO dashboard to control and visualize an AI agent playing Mario.

It started as a small AI experiment. Now it’s a full command center:

📊 Live Reward Graphs
🧠 AI Thought Feed
🏆 Achievement Toasts
📈 Level Completion Tracking
📦 Exportable Gameplay Sessions

🔧 What Powers It?

Python – AI loop, control logic, reward engine
RetroArch – NES emulator backend
Flask + Socket.IO – Live dashboard
Plotly.js – In-browser visual stats
JSON Logs – All gameplay and AI memory
Local AI – (DeepSeek via Ollama or Transformers if desired)

📺 Key Features

✅ AI State + Reward Dashboard

See Mario’s:

Live screen
Current decision
Score over time
Frame-by-frame commentary

🛠️ Full AI Control from Your Browser

Pause the AI. Reset the game. Log an episode. Save or load savestates — all without touching your terminal.

🏆 Toast-Powered Achievements

Real-time popups for:

🍄 Mushroom collected
🥇 Fire flower
🏁 Level completion
☠️ Death & 🔁 Retry
🏅 New high score
🏃 Longest survival streak

📈 Stats & Graphs

Auto-generated graphs show:

Level attempts vs completions
Reward trends
Survival history (coming soon)

📤 Session Export

Download a full ZIP with:

Screenshots
Reward logs
Chat log
AI stats
Achievements

📁 GitHub Structure

game_memory/
  logs/
    frame_###.png       # Screenshots
    chat_log.txt        # Live thought log
    achievements.json   # Unlocked trophies
    ai_stats.json       # Score & survival records
    level_stats.json    # Attempts & completions
utils/
  reward_dashboard_live.py  # Full Flask dashboard
main.py                     # Core AI gameplay loop

🧠 How the AI Thinks

No vector store is required — but you could add one.

Currently, the AI:

Reads RAM from RetroArch saves
Describes state in natural language
Logs rewards and performance
Optionally uses prompts or LLMs to generate input

You can enhance this with RAG, fine-tuning, or real-time retrieval.

🚀 Try It Out

Clone the repo: bashCopyEditgit clone https://github.com/CptNope/game_ai_rag_reproach_livechat cd game_ai_rag_reproach_livechat
Install dependencies: bashCopyEditpip install flask flask-socketio plotly pillow
Launch the dashboard: bashCopyEditpython utils/reward_dashboard_live.py
Run the AI player: bashCopyEditpython main.py

Visit http://localhost:5000 and watch your AI go!

📚 What’s Next?

Add real-time training visualization
Support multiple savestate slots
Serve frame stream as MJPEG video
Connect to a vector store (Reproach, ChromaDB, FAISS)
Enable fine-tuning via JSONL logs and LoRA

👾 Why I Built It

I love merging AI with retro games. There’s something special about watching an old classic teach a modern model. This project blends:

Game design
Machine learning
Interface dev
Real-time systems

…and it’s just the beginning.

🔗 Check it out on GitHub
📢 Feel free to fork, remix, or hit me up with your mods!

JeremyAnderson.Tech