AI & Agents for Engineering

Engineering Intelligence, Not Just Automation

AI at Hrulabs is not about chatbots or content generation. It is about applying large language models, autonomous agents, and computational reasoning to hard engineering problems — geometry generation, simulation pipelines, design space exploration, and extracting usable data from scientific literature.

Computational Engineering as Code

The most consequential shift in engineering today is not AI replacing engineers. It is engineers encoding domain knowledge as software, then systematically exploring design spaces that would take years to cover by hand.

The idea is simple: rather than drawing geometry in CAD, you write a codebase that encodes physics, manufacturing constraints, and material behaviour, and the code generates the geometry. Change a parameter and the design updates. Add a constraint and the model adapts.

This is the direction Hrulabs is pursuing. For parts governed by rules — heat exchangers, ducting and manifolds, enclosures, print fixtures — we build models where:

• Geometry is generated, not drawn — parametric Python produces STEP/STL output driven by constraints
• Simulation sits inside the design loop — analysis results feed back into geometry regeneration
• Manufacturing rules are first-class — print orientation, support requirements, and tolerance stacks are constraints in the model, not afterthoughts

What We Build

Code-Driven Geometry Generation

Parametric geometry generators in OpenSCAD and Python using the FreeCAD API that produce manufacturable 3D geometry from engineering requirements, rather than from a modelling session someone has to repeat by hand.

Simulation Pipeline Automation

Python orchestration layers that drive parametric sweeps — FreeCAD FEM or our own numerical codes — collect results, and generate structured reports without manual intervention between runs.

Autonomous Research Agents

LLM-powered agents that read scientific papers, extract structured data — material properties, process parameters, experimental results — and populate design databases. The aim is to cut the manual extraction work in a literature review, with a person still checking what the agent pulls out.

LLM-Augmented Design Review

Structured pipelines that ingest simulation results, dimensional inspection reports, and test data, then surface anomalies, flag constraint violations, and suggest parameter adjustments.

Custom Agent Frameworks

Purpose-built multi-agent systems using the Anthropic Claude API and the Claude Agent SDK for tasks requiring planning, tool use, and iterative reasoning across long engineering workflows.

Where It Works, and Where It Doesn't

Code-driven design pays off when a domain is physics-dominated, well-understood enough to express as constraints, and revisited often enough to amortise the model. The first version of any such model is more expensive than simply drawing the part — it only wins from the third project onward.

It works less well for one-off geometry, strongly aesthetic design, and topology-optimised structures with non-obvious connectivity. We will tell you which case you are in before taking the work.

Deliverables

• Parametric geometry source code, documented and runnable by your team
• Generated CAD output (STEP/STL) with the parameter set that produced it
• Simulation automation scripts and structured result sets
• Agent/pipeline source with prompts, tool definitions, and evaluation notes
• Handover documentation covering assumptions, valid parameter ranges, and known limits

Technologies