FutureCraft

Designing the future of flight at the speed of thought.

Comment

💡 Inspiration

Aerospace engineering is traditionally a slow, mathematically intense process. Engineers spend hours in complex CAD software just to visualize a basic concept, only to find out later that the physics doesn't work.

  • We asked ourselves: What if designing a plane was as fast as describing it?

We wanted to build a tool that bridges the gap between imagination and engineering. Not just generating pretty 3D pictures, but creating validated, physics-aware engineering models in real-time. We wanted to move from "drawing" to "engineering at the speed of thought."

✈️ What it does

AeroCraft is a real-time AI engineering assistant. It allows users to design aircraft using natural language, but it adds a layer of rigorous engineering physics on top.

You type "Create a high-altitude glider with carbon fiber wings," and the 3D model appears instantly.

Material Science: The user can select real-world materials (Titanium, Al7075, Carbon Fiber). The system calculates the aircraft's weight and structural integrity based on density 𝜌 ρ and tensile strength.

Physics Simulation: Unlike standard 3D generators, AeroCraft runs real-time beam theory calculations. It determines if your wing aspect ratio is aerodynamically viable and calculates safety factors against structural failure.

AI Optimization: If your design is unsafe (e.g., wings too long for the material strength), the AI detects the physics failure and suggests corrections.

⚙️ How we built it

We focused on Ultra-Low Latency as our primary metric. An engineering tool cannot feel "laggy."

The Brain (Cerebras): We integrated Cerebras Inference (Llama 3.1-70b). Standard LLMs took 3-5 seconds to generate complex JSON parameters. Cerebras does it in milliseconds. This speed allowed us to create a feedback loop where the AI can "check its work" against the physics engine instantly.

The Backend (FastAPI): We built a Python backend that handles the heavy lifting—physics equations, geometry validation, and API orchestration.

The Physics Engine: We implemented structural analysis algorithms to calculate stress and deflection, ensuring the models aren't just visual, but viable.

Infrastructure & Deployment:

We designed the architecture to leverage LiquidMetal Raindrop (SmartBuckets) for secure asset storage and Vultr Cloud Compute for scalability.

For the hackathon demonstration, we containerized the entire application using Docker and deployed it to Hugging Face Spaces. This ensured a stable, live prototype that judges can test immediately, regardless of regional infrastructure availability.

Challenges we ran into

  1. The "Hallucination" Problem: Getting an LLM to output perfect 3D geometry coordinates is difficult. The model often wanted to create wings that were disconnected from the fuselage. We solved this by using System Prompt Engineering with strict JSON schemas and a validation layer that "snaps" components together before rendering.

  2. Infrastructure Blockers: We faced significant hurdles with cloud provisioning. Despite having credits, regional payment gateway restrictions prevented us from activating Vultr instances, and we encountered Windows-specific compatibility bugs with the Raindrop CLI.

The Pivot: Instead of giving up, we adopted a Containerized Microservice approach. By wrapping our app in Docker, we decoupled it from specific cloud providers, allowing us to ship a working product on alternative hosting while keeping the Raindrop integration code ready for future deployment.

Accomplishments that we're proud of

We are most proud of transforming a basic visualizer into a simulation tool. We significantly extended the open-source base we started with:

  • Static Mesh Real-Time Physics Engine -(Stress, Strain, Aerodynamics)
  • Speed Standard Inference Ultra-Low Latency via Cerebras integration
  • Materials None Material Database (Titanium, Carbon Fiber, Aluminum)
  • Intelligence Text-to-Shape Self-Correcting Engineering Agent ## 🧠 What we learned

Latency is User Experience: In creative tools, the difference between 0.5s (Cerebras) and 4s (Standard LLM) is the difference between "flow state" and frustration.

Resilience in Deployment: We learned the value of Docker. When one cloud door closes (due to verification errors), containerization opens another.

The Complexity of 3D: We gained a deep appreciation for the math required to translate text into Three.js geometries.

🔮 What's next for AeroCraft

CFD Integration: Moving from basic beam theory to simple Computational Fluid Dynamics for wind tunnel simulation.

Full Cloud Integration: Resolving the account verification steps to fully migrate the storage layer to Raindrop SmartBuckets and scale the compute on Vultr.

Export to STL: Allowing users to 3D print their generated drones and planes.

Built With

Share this project:

Updates