Interview
It asks about loads, how it mounts, the space it has to fit, quantity and process. Only what changes the answer. Then it writes it all down for you to check.
Agentic CAD design, proven and documented
You describe the job. LMCAD designs the parts, puts the assembly together, proves it holds under load, checks a shop can actually make it, and hands you the whole package: drawings, parts list, inspection sheets and the report behind them. If you want to know where a number came from, you can go and look.
What is LMCAD
LMCAD is AI CAD software for mechanical engineers. You write down what a part has to do, and it handles the rest: the geometry, the finite element analysis that proves it holds, manufacturability checks against your process, and the drawings, bill of materials and engineering report at the end. Design automation for the ordinary parts that never get looked at properly, running on your own computer.
You size a part. You hand it over. You wait. Days later you find out it was too thin, and you start again. Every one of those gaps is a person waiting on another person. LMCAD closes the whole design and simulation loop on its own, iterating the CAD geometry against your requirements as many times as it takes.
It asks about loads, how it mounts, the space it has to fit, quantity and process. Only what changes the answer. Then it writes it all down for you to check.
Beam theory, thermal resistance, pressure drop. Seconds, not minutes. A design that fails here never reaches a solver.
Parametric geometry with named faces, standard fasteners from a library, and manufacturability checked against the process you named.
Mesh, solve, extract. Convergence is demonstrated rather than assumed, and the result is cross-checked against the estimate.
Every requirement gets a verdict: pass, fail, or not evaluated. Nothing is called done on a requirement nobody measured.
It moves declared parameters inside bounds you set and runs again, or proves the requirements conflict and says which to relax.
Dimensioned sheet, BOM, STEP file and a report where every figure links to the run that produced it.
Each judged iteration is a version. Compare two and see mass, safety factor and which verdicts flipped. Not a text diff.
You are the one signing the drawing, so you need to see the working. Every number arrives with the run that produced it, the assumptions it rests on, and an honest note wherever something could not be answered.
$ lmcad run brief motor-bracket rev 3 · 4 requirements material Al 6061-T6 E 68900 MPa σy 276 MPa estimate cantilever theory → t ≥ 7.2 mm iteration 1 t 8.0 ribs 3 mesh 41 208 nodes quadratic quality ok solve static 4.1 s check FEA 0.541 mm vs theory 0.529 mm 2.3% ok σ max 289 MPa at rib root → SF 0.95 FAIL R1 iteration 4 t 10.0 ribs 4 σ max 118 MPa → SF 2.34 PASS R1 mode 1 341 Hz PASS R2 mass 212 g PASS R3 R4 fatigue not evaluated: brief defines no load spectrum. released as unverified pending R4
| ID | Requirement | Result | Verdict |
|---|---|---|---|
| R1 | Yield safety factor ≥ 2.0 | 2.34 | Pass |
| R2 | First natural mode ≥ 300 Hz | 341 Hz | Pass |
| R3 | Mass budget ≤ 250 g | 212 g | Pass |
| R4 | Fatigue life ≥ 10⁷ cycles | n/a | Not evaluated |
Nobody told it how often this part gets loaded, so it cannot answer the fatigue question and it says so on the front page of the report. A missing answer is something you can act on. A made-up one ends up in production.
Every solve is compared to an independently derived closed-form estimate. An order-of-magnitude disagreement halts the run, because the model is broken, not the physics.
Peak stress at a sharp internal corner rises forever as you refine. The system knows the difference between a converging result and a singularity, and refuses to report the second as a number.
A design from last year rebuilds exactly, because every project records the conditions it was produced under. If anything has shifted since, you are told rather than quietly handed a different part.
You do not set up the simulation. You ask the question. It picks the right analysis, sets the boundary conditions, meshes the part, and decides whether the answer can be trusted. Finite element analysis without needing an FEA specialist in the building.
Stress, deflection and safety factor under real load cases. Plasticity, contact between parts, large deflection, bolt preload and weight optimisation where the load path needs them.
Natural frequencies and mode shapes clear of your excitation, harmonic response, impact, and buckling for anything slender or thin-walled. Where a structure radiates noise, the vibration and the sound it makes are solved together.
Steady and transient heat transfer, conduction, convection and radiation, and coupled thermal-structural analysis for parts that grow as they warm.
Internal flow and pressure drop, external aerodynamics with drag and lift, fans and rotating machinery, conjugate cooling, and shape optimisation driven by the flow result.
Motor torque, cogging, back-EMF and losses, magnetic circuits, induction heating, and the coupled thermal picture that comes with them.
Full range-of-motion sweeps with interference detection and swept envelopes, and the joint reactions that come out feed straight back in as load cases.
Real work is rarely one bracket on its own. LMCAD does full assembly design: it models the parts, works out how they meet, and holds every joint fixed while each piece is worked on separately. Which is why they still fit when it puts them back together.
Every joint is an agreed contract: mating faces, bolt pattern, fits and the load crossing it. Change one side and the other is told.
Static clearance across the assembly, and through the full range of motion for anything that moves. Tolerance stack-ups on the chains that matter.
Bill of materials rolled up through every sub-assembly, with standard fasteners, materials, quantities and an estimated cost per unit.
An ordered build sequence with exploded views, torque values and the fasteners needed at each step, in a document someone on the floor can follow.
If a machinist has to ring you to ask what something means, the drawing failed. So nothing goes out until every field is filled in, every dimension points at something real, and no two views sit on top of each other.
The definitive geometry in the neutral format your shop already programs from, part by part and as the built assembly.
Views and sections, hole and thread callouts taken from the model itself, fits resolved to limits, tolerance notes and a revision block.
Exploded views, balloon references to the parts list, section views through the joints, and the fasteners called out where they go.
Rolled up through every level with quantity, material, specification, standard-part references and estimated cost.
Ballooned drawings and a measurement table, with the characteristics the analysis actually depended on marked critical.
An ordered sequence with exploded figures, fasteners and torque values, written for the person doing the work.
Requirements verdict table, the analysis behind each one with plots, every assumption made and why, and provenance for every figure.
The system verifies. A person releases. Acceptance is recorded against the revision and the exact geometry it was given.
The people who get the most out of this are not the ones with a simulation department. They are the ones who have always known a part should be checked properly and have never had the hours. One engineer, a queue of parts, and everything due at once.
Special machines are mostly one-off parts that still have to hold. Get them sized, checked and drawn while you are working on the rest of the build.
Every figure traces to the run behind it, inspection sheets come with the drawing, and the whole thing runs inside your own network with nothing going out.
A customer sends a rough idea. Turn it into a proper package with a parts list and a cost, fast enough to win the job.
The unglamorous parts nobody has time to analyse are the ones that come back. Size them properly without slowing the programme down.
If your whole mechanical team is two people, this is how they cover the work of five without any of them working a weekend.
Say yes to the next project before you have found the person to staff it, and hand the client a package that looks like a bigger firm made it.
A seat per engineer, billed annually. Every plan has every kind of analysis in it, so you are never told a capability costs extra. What changes as you go up is how your team shares seats, how fast we answer, and how much of the setup we do alongside you.
In every plan, without exception
For the engineer who has been meaning to check these parts properly for years.
For a team that wants the whole group working the same way.
For regulated work, or anywhere the network does not reach.
Every plan starts with a paid pilot on a part you already have drawings for. If it cannot match what you know to be true, you do not go ahead and you owe nothing further.
Per seat, per month, billed annually in Indian rupees, exclusive of taxes. Month to month is available at a 25% premium. No setup fee, no charge per analysis, no charge per drawing.
Everything the system produces belongs to you, in open formats, on your own disk. If you stop paying, the files do not stop working.
It takes a written description of what a part has to do and carries the job through to a package you can send out. It designs the geometry, sizes it, runs the analysis that shows whether it holds, checks a shop can make it, and produces the drawings, bill of materials and report. Every figure in that report links back to the run that produced it, so the engineer signing it can see the working rather than take it on trust.
No, and it is not built to. It does the hours of work that sit between having an idea and having something you can release. Deciding whether the requirements were the right ones, whether the assumptions hold for your application, and whether the part is fit to ship are judgements that carry responsibility, and responsibility is not something software can hold. Nothing leaves as released until a person signs it.
Three things, and you can check all of them. Before any simulation runs, the part is sized by hand calculation from published formulas; if the simulation later disagrees with that estimate by an order of magnitude, the run stops, because a gap that size means the model is wrong rather than the theory. Mesh convergence is demonstrated by solving at several densities and showing the answer settling, instead of reporting one mesh and hoping. And where a result cannot converge, such as the stress at a perfectly sharp internal corner, it is reported as that rather than as a number.
No to both. The software, the analysis and your files all run on your own computer. We hold your licence and how much of your allowance you have used, and nothing else. Since your geometry and results are never uploaded, there is nothing of yours here to train on even if we wanted to. If you connect your own model provider account, the written instructions go straight from your machine to that provider without passing through us.
Assemblies. It designs the parts, decides how they meet, and fixes each joint as a contract before the individual parts are worked on, which is why they still fit when it puts them together. It checks clearance across the assembly and through the full range of motion for anything that moves, rolls the bill of materials up through every sub-assembly, and produces assembly drawings with exploded views and a build sequence.
A STEP file of the model and the assembly, dimensioned drawings as PDF and DXF, the bill of materials, an inspection sheet with the measured characteristics balloon-referenced, and the engineering report. Those are the formats every CAD and CAM system reads, so you can open a part in SolidWorks, Fusion or Creo, change it and carry on. The files are yours and they keep working whether or not you keep paying us.
It tells you, and it tells you which one to relax and roughly by how much. Proving that a set of requirements conflict is a real engineering result, and a more useful one than a design that quietly satisfies three of them and ignores the fourth. Anything it could not evaluate at all is marked not evaluated and appears at the front of the report, never counted as a pass.
The solving methods are the established ones, so on a problem both can handle, the numbers agree closely. The difference is not the solver, it is the scope. Those tools expect you to arrive with a finished model, know which analysis to run and how to set it up, and then take the result away to draw. LMCAD does the design, decides the analysis, judges whether the answer can be trusted, and produces the paperwork. It is not a faster solver. It is the work around the solver.
The only demo worth your time is one where you already know the right answer. Pick a part you have drawings for, tell us what it has to survive, and we will run it while you watch. If it gets it wrong, you will see that too.