The Logical Path Narrator

Project Overview

The Logical Path Narrator is a first-person puzzle game developed over a few days for the Brackeys Game Jam 2026.2, under the theme “Trust no one”. You awaken in a mysterious, isolated facility with only one companion: a dynamic, ever-watchful pathological AI narrator that monitors your every move, contextualizes your interactions, and delivers sarcastic real-time commentary.

While the jam rules strictly prohibited “AI-generated content” (pre-generated assets), I built this game to test a different concept: integrating a live, real-time Large Language Model (LLM) API directly into the gameplay loop. Players who want a strict adherence to the jam rules can simply input an invalid API key in the main menu to break the narrator and play the physics puzzles in silence.

Rapid Skill Acquisition: Learning Unreal Engine

This project was built from scratch in under 3 days and marks my very first time using Unreal Engine. As a developer looking to shift toward Forward Deployed Engineering, I used this opportunity as an intensive, deep-dive learning experience to prove my ability to rapidly acquire new technical stacks and integrate complex APIs into unfamiliar environments.

Working entirely within Unreal’s Blueprint visual scripting system, I successfully learned the engine’s core systems and implemented a fully functional game loop, including:

• Core Game Systems: A smooth first-person controller, enhanced input systems, and persistent save game data structures.
• Physics Interactions: Satisfying mechanics allowing the player to pick up, carry, push, and toss objects to solve environmental puzzles.
• State Management: An alternative to the singleton pattern using the GameInstance class to handle global events and API states across level transitions.
• Dynamic UI & Menus: A fully functional main menu with graphics settings and customizable API configuration.

Technical Architecture: Real-Time AI Integration

The standout feature of the game is the pipeline I constructed to feed gameplay context to an LLM and parse the response back into the world as a dynamic narrator.

• Context Gathering: I designed a system of spatial trigger volumes and interaction event dispatchers. When a player enters a specific room, drops a box, or stares at a wall too long, the game constructs a contextual prompt behind the scenes.
• HTTP API Integration: I built an asynchronous HTTP request pipeline to communicate with standard OpenAI-compatible APIs (supporting providers like OpenRouter or Google Developer Console). The game ships with my free-tier key (I know this is not the best practice, but this is a jam), but allows players to paste their own into the UI.
• Local Text-to-Speech: To bring the AI’s sarcastic responses to life, I integrated Piper TTS. This required upgrading specific .NET frameworks to compile a C++ plugin so it would run locally on the user’s PC. However, this offline generation unfortunately takes a few seconds too long for snappy real-time gameplay and I was not able to make the streaming work in the limited time available (prompting me to rely on native OS TTS for the default rapid responses in the final build), though it remains available as a toggle in the settings.

Expanding Academic Interests

My fascination with integrating generative AI into interactive environments extends beyond this jam. This project serves as a practical, hands-on continuation of the academic research I supervise at Masaryk University. Specifically, it builds upon the concepts explored in A Quest for Information: Enhancing Game-Based Learning with LLM-Driven NPCs, where we researched how LLMs can generate dynamic, context-aware narratives and even interact back with the game to enhance user engagement.

Building The Logical Path Narrator allowed me to transition those theoretical design concepts into a functional, deployed software pipeline under extreme time constraints.