SUAKIN
Toroidal Designer
About this API
An engineering intelligence layer over simulation for UAV design, starting with toroidal propellers for quadrotors. It is not a simulation tool, not an LLM wrapper and not a CAD frontend. It is a design engine with memory, governance and reproducibility guarantees, delivered as an API.
Who it is for
Solo engineers, UAV startup teams and small drone companies. These are people who need a disciplined, reproducible workflow without enterprise bureaucracy. Structural and small aerospace R&D teams are a later audience.
What works today
This is the honest list. Everything here is built and in use.
- BuiltToroidal propeller geometry generation. Closed-loop B-rep, exported as STEP in four progressive stages (rails, sweep, fused, validated).
- BuiltB-rep-only validity: OCP topology, watertight shells, and a Euler-Poincaré genus gate. No tessellation, because a mesh can read watertight while the underlying B-rep is not.
- BuiltGraph-driven job pipeline across control / data / execute blocks, with surfaces as contract injectors and job state in Firestore.
- BuiltStation-count policy, served at
GET /policies/toroidal_stations. - BuiltAerodynamic feasibility check (M1 to M6) at
POST /validate/design. It is informational and does not gate generation.
The MVP does one thing
Generate a valid toroidal propeller B-rep and emit it as STEP. Everything below the geometry line is deliberately deferred until that is solid: the open-blade comparator, CFD evaluation, thrust and efficiency metrics, acoustic proxies, surrogate models and optimization loops.
Principles the system is built on
- Determinism over convenience. No hidden defaults, no silent parameter drift. If it cannot be reproduced, it is a bug.
- Auditability over speed. Every run is meant to be traceable to its exact inputs, versions and outputs.
- Policy over improvisation. Schemas define structure; policies define allowable behaviour.
- Human approval over autonomy. An LLM may propose a plan or read metrics; it may not execute compute, bypass policy or override an approval.
- Multi-scenario honesty. Evaluations must never collapse into a single opaque number. Per-scenario metrics are preserved, and the aggregation method is declared.
What is planned, not built
Stated plainly so nobody plans around something that does not exist yet.
- PlannedRun manifests. Every job emitting its exact inputs, policy bindings, solver version, container hash, seeds and output hashes. This is the contract of truth for reproducibility.
- PlannedPolicy layer. Versioned solver, mesh, compute, design and objective policies, with approval gates and budget limits. Today the registry holds station counts only.
- PlannedCFD evaluation. RANS-based thrust, torque and efficiency, plus an acoustic proxy for the toroidal-vs-baseline noise comparison that motivates the geometry.
- PlannedSurrogate models and bounded optimization, with CFD confirmations limited by an explicit compute budget.
- PlannedMulti-tenancy. Tenant and project scoping on every artifact, with no cross-tenant joins or shared retrieval memory.
- PlannedOpen-blade comparator for like-for-like toroidal vs conventional studies.
Further out
Frame structural integrity, payload integration, energy modelling and full mission-envelope optimization, with design-to-manufacture constraints for toroidal printability throughout. The destination is a governed engineering operating layer for UAV system design, combining physics-based simulation, learned models and controlled optimization under explicit policy and cost bounds.
What it refuses to be
A black-box optimizer, a generic CFD hosting service, a chat wrapper for engineering, an uncontrolled auto-design generator, or a no-code physics sandbox. Engineering credibility is mandatory.