I integrated a local Llama 3.2 model to act as a dynamic Dungeon Master in my indie RPG.

Hey everyone, I am not trying to sell or self promote mainly just wanted to showcase a big project I've been working on ever since I started studying data science and artificial intelligence and integrating AI into workflows and using it as an augment to create things that were previously out of reach for so many people, because if used right it can become a second brain and not a crutch. I’m the solo dev behind Void Runner , an isometric ARPG/MOBA hybrid built in Python. I recently hit a wall with traditional procedural quest generation. Hand-crafting templates gets repetitive fast, and players quickly learn the patterns to these things whether you like it or not. To solve this, I built the "Void Caller AI" , a system that uses a local, quantized Llama 3.2 model to act as a dynamic Dungeon Master. Instead of just generating random flavor text, the system uses a lightweight RAG (Retrieval-Augmented Generation) pipeline. It reads live server telemetry (who died, what items were looted, which bosses were defeated recently) and weaves those actual server events into the narrative of the quests it generates. Because it runs locally via Ollama on our backend, there are no crazy cloud API costs, and latency is kept completely manageable. Here is a simplified look at how the Python backend bridges the SQLite telemetry with the Llama 3.2 prompt: import json import ollama from sqlalchemy import text from database import SessionLocal def generate_dynamic_quest(difficulty: str, target: str): db = SessionLocal() # 1. Fetch recent server telemetry for context (RAG-lite) lore_context = "" try: # Grab recent server events to weave into the narrative recent_events = db.execute(text( "SELECT username, event_type, dungeon_type FROM ai_events ORDER BY id DESC LIMIT 3" )).fetchall() if recent_events: events_str = "; ".join([f"Runner '{r[0]}' triggered a '{r[1]}' in '{r[2]}'" for r in recent_events]) lore_context = f" Incorporate this recent live server telemetry into the lore: {events_