Back to Work
02AI / Gaming

Generating Playable Games with AI

An AI-powered game prototyping platform that turns a text prompt into a playable build using LLM pipelines, RAG over engine docs, and automated asset generation.

Generating Playable Games with AI interface
2023 — 2024
Solo
10 months
React · TypeScript · Python

The Problem

Prototyping a game idea takes weeks of engineering and art work before anyone can tell whether it is fun.

  • Slow idea-to-prototype loop
  • Art and code blocked on each other
  • Hard to test many mechanics quickly
  • Prototypes rarely reach a playable state

Our Solution

A generation pipeline that composes game logic, scenes, and assets into a runnable build that designers can play and iterate on in minutes.

  • Prompt-to-playable build pipeline
  • RAG grounded in engine documentation
  • Automated sprite and audio generation
  • In-browser preview and versioning

Architecture Overview

Key Engineering Challenges

Deterministic Output

Making generation reproducible enough that designers could iterate rather than reroll.

Build Reliability

Auto-repairing generated code so a broken build never reached the playtester.

Asset Coherence

Keeping generated art consistent in style across an entire scene.

Cost Control

Caching and routing between models to keep per-generation cost predictable.

Tech Stack

Frontend

ReactTypeScriptCanvasZustand

Backend

PythonFastAPIQueue WorkersSandboxing

AI

LLM PipelinesRAGDiffusion ModelsEvals

Infra

DockerAWSObject StorageObservability

Impact & Results

10 min

Idea to playable build

85%

Builds runnable first try

6x

Faster prototyping

200+

Prototypes generated

We tested more mechanics in a week than we used to in a quarter.
Game Designer

What I'd Improve Today

Evaluation Harness

Automated playability scoring so weak generations never reach a human.

Asset Style Locking

Train a lightweight adapter per project to lock visual identity.

Multiplayer Templates

Extend the pipeline to networked prototypes, not just single-player loops.

Constrain the generation space before improving the model.

Auto-repair loops matter more than raw model quality.

Designers judge prototypes by feel, so latency is a feature.

Interested in building something similar?

I'm open to exciting products, collaborations, and engineering opportunities.

Let's Connect