Concept Forge
By Kshitij Aucharmal • 4 minutes read •
ConceptForge: Building My AI-Powered Simulation Engine from Scratch
“Describe your world in English. Let the AI bring it to life.”
What I’m Building
Over the next six months, I’m working on ConceptForge — an AI-powered simulation engine built completely from scratch, using Vulkan, C++, and a layer of natural language processing magic.
This isn’t just a simulation engine. It’s the start of something more — maybe even a game engine down the line. Here’s the dream:
- A low-level, high-performance rendering and simulation engine in C++, powered by Vulkan.
- Physics, real-time simulation, ray tracing (yes, built myself — no Vulkan RTX extensions).
- A Python API using pybind11, so users can write high-level logic easily.
- LLM integration — users can describe scenes in plain English, and the AI will generate Python code behind the scenes.
- GUI using GLFW and ImGui to provide visual feedback, simulation controls, and editing tools.
All of this is open-source and evolving here:
GitHub: ConceptForge
High-Level Architecture
This is the general structure I’m working with: 
It’s a full-stack creative pipeline — except instead of websites, it makes interactive 3D simulations based on natural language prompts.
Tech Stack
| Tech | Purpose |
|---|---|
| C++ | Core simulation logic, ECS, physics, and rendering backend |
| Vulkan | Raw rendering control — full GPU access, custom pipelines, ray tracing |
| GLFW | Window/context management |
| ImGui | Editor UI, debug controls, interactive tweaks |
| pybind11 | Binding layer to expose C++ functionality to Python |
| Python | User scripting interface (and what the AI generates) |
| LLMs | Converts English into code using GPT or custom small models |
Vulkan in Simple Terms
Why Vulkan? Because I want total control over how things render, move, and interact — right down to how memory is allocated on the GPU.
Think of Vulkan like this:
- OpenGL holds your hand. Vulkan tells you to build the chair you’re gonna sit on.
- That means more boilerplate, but also massive control and zero hidden costs.
- Perfect for things like: custom ray tracing, compute shaders, massive simulations, and crazy GPU tricks.
The Vision
Here’s the core loop I’m building toward:
User: “Spawn a red sphere on a green plane. Drop a blue cube from 5 meters with gravity.”
- The LLM parses this text and generates corresponding Python code using the engine’s API.
- Python runs →
pybind11passes it to the C++ backend. - Simulation is updated, rendered using Vulkan, displayed in the GLFW window.
- User can interact, tweak parameters, or add more prompts.
This bridges natural language + code + simulation, and gives users an almost magical interface to create with.
What You’ll Be Able To Do
Eventually, this will support things like:
- Prototyping with AI assistance
- Simulating complex systems: physics, particles, ray tracing, ray marching, etc.
- Procedural scene generation via LLMs
- AI agents interacting in a simulated world
- Realtime interactive visualization for data or experiments
Visual Overview
Here’s a high-level diagram of how it all connects:
Why I’m Doing This
I’ve always wanted an engine where I could just describe what I want, and it would spring to life. We’re at a point now where LLMs are good enough to bridge that gap — but most engines are built from the top-down, and not designed for this kind of integration.
So I’m doing it differently: bottom-up, from GPU code to English-language scripting, with a clean and extensible architecture in mind.
This is going to take time, but I’m documenting every step, and you can follow the journey (and maybe even contribute) on GitHub.
The real reason can just be cause I want to :P
Follow the Project
Blog posts, demos, and releases will be linked here as they come out.
Got ideas? Want to contribute? Drop me a message or a GitHub issue. Let’s build something wild. 🔥