Physics
Derives
Geometry
One plain-english sentence becomes a STEP file, FEA report, and auditable engineering package. No CAD. No hallucination. No cloud.
Computed Values
Shigley Eq 6-2
AI handles ambiguity. Math handles precision.
Language models are excellent at understanding intent. They are terrible at engineering precision. Rigel uses each for what it does best, with an absolute boundary between them.
AI Touchpoint 1
01NLP Parser
Parses plain text into structured engineering parameters. Material, load magnitude, load direction, hole pattern. Returns null on uncertainty, never guesses.
AI Touchpoint 2
02Result Interpreter
Translates FEA output into plain English. Flags anomalies for engineer review. Never omits stress concentrations or recommends design changes.
AI Touchpoint 3
03Suggestion Engine
Proposes parameter changes with reasoning only. Requires human approval before anything runs. Never bypasses the engineer.
Shigley's Eq 6-2: Flange Thickness from Bending Stress
Live Equation
How it actually works
Describe what you need in plain English. The kernel computes dimensions from Shigley's equations. Four auditable files land in your output folder.
Spec Ingestion
LLaMA 3.2 3B extracts material, load magnitude, load direction, and hole pattern locally via Ollama. If any field is ambiguous or missing, the pipeline halts and asks. It never guesses. It never defaults silently.
Physics Solver
Flange thickness from bending stress. Web thickness from shear. Fillet radius from stress concentration target. Zero hardcoded values in the geometry kernel. Every number traceable to an equation and section number.
FEA Verification
Gmsh generates a Tet10 volume mesh. CalculiX runs the static step. Convergence is confirmed when stress changes less than 5% between the two finest meshes. Von Mises stress and safety factor parsed from the .frd output.
Performance Metric
8–12 minutes (8-core laptop)
“On a standard 8-core laptop, the full pipeline currently takes 8–12 minutes. We measured it. We are not claiming 5 minutes.”
Six modules. Zero black boxes.
Each module maps to a stage in the aerospace analysis workflow. From natural language input to auditable FEA results. No closed solvers, no opaque defaults, no guessing.
Constraint Parser
Parses natural language into structured engineering parameters. Material grade, load vectors, fastener patterns, and geometric constraints extracted in a single pass.
Analytical Kernel
Shigley's mechanical design equations implemented with aerospace-grade rigor. Every dimension derived from first principles, not interpolation or training data.
Geometry Kernel
Generates B-REP solids from analytical dimensions. Parametric modeling at the fidelity of professional CAD kernels used in turbine and airframe design.
Mesh Generator
Tet10 volume meshing with adaptive refinement at fillet radii and stress concentrations. Automated meshing that rivals commercial pre-processors.
Structural Solver
Static nonlinear solver with convergence verification. Von Mises and principal stress fields delivered at the same fidelity aerospace sign-off procedures require.
Audit Chain
SHA-256 hash chain linking every solver state, mesh snapshot, and design decision. The immutable computation record that DO-178C and AS9100 workflows demand.
Local-first
127.0.0.1
Runs on your machine, not ours
Audit-linked
SHA-256
Immutable computation record
Reproducible
Deterministic
Same input always yields same output
ANSYS Mechanical
$30k+/yr license, closed solver, opaque defaults
Analytical Kernel
Open equations, traceable to Shigley section numbers
CATIA / NX
Manual parametric modeling, human geometry decisions
Geometry Kernel
B-REP from physics, zero human dimensioning
HyperMesh + Radioss
Manual meshing pipeline, hours of setup per model
Mesh + Solver
Automated Tet10 meshing, CalculiX solve, < 12 min
Engineering Sovereignty
A physics-first model designed and hot-fired a 20 kN methalox engine in weeks. No CAD. No human geometry decisions. We are encoding that philosophy for Indian aerospace.
A physics-first computational model designed and hot-fired a 20 kN methalox rocket engine in weeks. No CAD. No human geometry decisions. Every dimension derived from encoded engineering knowledge.
Encoding that philosophy for Indian aerospace, starting with the smallest defensible domain. Static brackets under axial load, validated against physical tests, built to meet the same rigor as certified analysis workflows.
Physics first
Geometry is derived from equations, never drawn by hand
No silent defaults
Null on uncertainty. Never guess, never interpolate
Auditable by construction
SHA-256 hash chain from input string to final mesh
M-series brackets, Al6061, static axial load, static FEA
Multi-load cases & multi-material validation
Dynamic load response & fatigue analysis
??? — Cannot plan from here
Four files. Every time.
Every run produces the same four files. Geometry you can machine. Reports you can audit. A hash chain you can verify. No surprises, no missing pieces.
bracket.step
Production GeometryPhysics-derived parametric geometry ready for CNC or additive manufacturing
report.pdf
Stress ReportVon Mises stress and factor of safety
bracket.dxf
CNC Ready2D base flange projection, ready for the CNC shop
audit.json
Tamper-EvidentSHA-256 hash chain from input string to final mesh
Every pipeline run produces exactly these four files. No more, no less. If any file is missing, the run is invalid.
We test against reality. Not textbooks.
Three brackets. Three physical tests. FEA predictions published before the press test. Error margins published after. No backfitting.
Real parts fail from stress concentrations, manufacturing defects, and assumptions your equations did not make. We are loading physical brackets to failure in a workshop press at PSG College of Engineering, Coimbatore. The failure loads go in this table. The error percentages are public. If the error is 30%, we publish 30%.
PSG College of Engineering
Coimbatore, Tamil Nadu
Workshop press
Physical loading to failure
Public error data
Published percentages, no hiding
| # | Bracket | Material | Predicted (N) | Actual (N) | Error % | Status |
|---|---|---|---|---|---|---|
| 01 | CH_BRACKET_V2.1 Channel bracket | Al6061-T6 | 8,420 | — | — | scheduled |
| 02 | L_BRACKET_V1.0 L-bracket | Al6061-T6 | 6,100 | — | — | scheduled |
| 03 | GUSSETED_V3.0 Gusseted bracket | Al6061-T6 | 12,400 | — | — | scheduled |
Physical test date: Week 5 of the build plan
FAILURES.md is a public file.
Every wrong result before it was fixed. Published openly. No selective reporting.
Honest scope
What Rigel does today, what it does not do, and why the boundary matters more than the ambition.
Not This
This
Stage 1 of 4. The other three cannot be planned yet. That is not a limitation, it is how engineering works.
Four people. One bracket. No fantasy.
Each role owns exactly one file. No role overlap, no ambiguity, no single point of failure. Four people in Coimbatore who assigned responsibilities before writing a line of code.
Physics Owner
Owns the physics equations
Geometry Owner
Owns the geometry kernel
FEA Pipeline
Owns the FEA pipeline (pair)
Stage 1 is not done until changing the load automatically changes the geometry and a physical bracket agrees with the FEA within 15%.
One person.
If you build hardware that flies, burns, or holds pressure, you need equations, not interpolation. Rigel is for the engineer who would rather trust a textbook than a language model.
A junior mechanical engineer who needs a defensible first-pass structural answer before a review meeting. Not CATIA. Not Ansys. Something that runs locally, shows its equations, and does not hallucinate geometry.
And anyone who thinks Indian aerospace deserves its own computational engineering infrastructure built honestly from first principles.
Local Only
No cloud. No lock-in.
Immutable Math
Equations derived from Shigley's.
Physically Validated
Tested to failure in a press.
Stay close.
We build in public. The commit history is the progress report. When Stage 1 passes all five gates, you will know before we announce it.
Use it when it ships
v0.1 ships when three physical brackets agree with the FEA within 15%. One email when that happens. Nothing else.
Multi-load cases & multi-material validation
Dynamic load response & fatigue analysis
First public release: three brackets validated
If the commit history goes dark for more than a week, something went wrong. We will say so publicly.