Local AILocal AI

An assistant that can open the model

We install the open-source AI model you choose on your own hardware, connect it to your solvers and set the limits it must work within. It runs the steps your engineers would run and writes the result back into your reporting format.

Your choice of model

Any open-weight model your policy allows, served from your own GPU server or workstation. Works fully offline.

Drives the solver

Tool connectors expose meshing, loads, supports, solving and result extraction as safe, named actions.

Knows your library

Retrieval over years of scripts, reports and standards, so answers cite your own prior work.

Leaves a trail

Every call is logged with inputs and outputs, ready for design review and quality audits.

assistant · on-premsession 0419

Run the modal check on bracket_rev4 and give me the first six modes.

→ import_geometry bracket_rev4.stp ok → named_selections 14 found ok → fixed_support NS_MOUNT_FACES ok → generate_mesh 96,118 elements quality gate passed → solve_modal 6 modes · 41 s ok

Mode 1	817.38 Hz
Mode 2	1,204.61 Hz
Mode 3	1,988.07 Hz
Mode 4	2,341.72 Hz
Mode 5	3,096.40 Hz
Mode 6	3,512.95 Hz

✓ Report written: bracket_rev4_modal.pptx (11 slides)

Model open-weight LLMNetwork offlineData egress 0 bytes

Example session for illustration. Mode 1 matches our published modal GUI demo.

Local AIModels & limits

Install the model you want. Set the rules it follows.

You decide which open-source model runs, what it may touch and when a human must approve. We size the hardware, install the model and wire it into your engineering tools.

Open-weight model families we deploy

LlamaGeneral assistant, long context
QwenStrong at code and tool use
MistralEfficient on a single GPU
GemmaCompact models for workstations
DeepSeekReasoning and scripting tasks
PhiSmall models for edge and laptops
Your ownFine-tuned on your reports and scripts

Model choice depends on your licence policy, hardware and language needs. We benchmark candidates on your own tasks before deployment.

Guardrails

Extra limits on top of the model

Allow-list of tools: the assistant can read results but only run solvers you approve
Read-only mode for sensitive projects and released designs
Human approval before any solve, file change or message leaves the team
Scope rules: answers only from your library, with cited sources
Per-project access, so one customer's data never reaches another
Full audit log of prompts, tool calls and outputs

Integration

Works with your LLM and machine-learning systems

Connect to existing LLM platforms and internal chat assistants
Surrogate models trained on simulation results for instant predictions
Machine-learning models for anomaly detection in test and field data
Feed results into your MLOps pipeline, data lake or dashboards
Standard tool protocols, so new solvers plug in without rework
Trigger AI steps from automation flows, Teams or PLM events

Start here

Which analysis do you repeat every week?

Sketch it as a flow in two minutes. We reply with a feasibility check and an estimate.

Build your flow Contact us