Research · Research
AeroForge
A fully local design pipeline in which AI agents derive an aircraft configuration from physics and task alone — and which refuted its own assumptions.
What emerges when you give an AI no shape, only physics and a task — offline, without cloud CAD and without data leaving the building? AeroForge is our experimental setup for this question. Not a product, not an aircraft: a pipeline that generates, checks and discards candidates.
The problem
Agentic systems sound impressive in presentations. Whether they produce anything usable under real constraints — weight, power, energy — only becomes apparent when a physics model calculates against them. That is exactly what we wanted to see.
How it works
- 01 A local knowledge store of datasheets, a biomimicry library and a parts catalogue gives the agents context.
- 02 A designer agent generates CAD code; a simulator executes it in isolation with a time limit; a physics critic approves it or sends it back into the correction loop.
- 03 An in-house physics model recalculates design points, flight time and efficiency — re-checkable, not asserted.
- 04 A local portal shows the 3D model, sections, live sliders and library; a chat against the local model can adjust the sliders via validated control commands.
Sovereignty & evidence
- Entirely local on a single machine; models via Ollama, knowledge store local, 3D rendering without third-party servers.
- An honest limit: the simulator is process isolation, not a security sandbox.
Why we show this
Because the strongest result is a negative finding. The pipeline took apart the figures from its own master plan and replaced them with re-checkable ones. That is exactly the property we want in every system we build for customers: better to refute an assumption than to carry it into production.
Status
Pipeline, portal and test suite are running; a target design has been chosen. Nothing has been built or flown. The project remains research.
Frequently asked questions
- Does the thing fly?
- No. There is no built device, no test-bench run, no flight. All flight-time and payload figures are model calculations and are explicitly labelled as such in the project.
- So what is the result?
- That the pipeline refuted its own initial figures: an assumed propulsion system actually needed three to six times the power, an assumed sensor does not exist in that weight class, an assumed battery had been calculated too light. A system that finds its own mistakes is worth more than one that delivers pretty figures.
- Does the AI design autonomously?
- No. The pipeline generates and checks candidates; the target design was chosen by a human. That is how it should stay.
A comparable challenge on your desk?
Send us three sentences. We reply within one business day with an honest assessment — including when we are not the right partner.