Challenge 2
Agentic AI for Automated CAD Generation and Autonomous Simulation _ SIAD Group
Team:
Building Blocks (lead) and MEP Engineering — consortium of Spanish SMEs
Proposal title:
SKIDGENT — Agentic AI for constraint-aware compressor-skid piping design and autonomous simulation
Organisation
Building Blocks is a Spanish software SME based in Elche, founded in 2022 and holding the “PYME Innovadora” seal, supporting R&D projects with AI, data and simulation technology. Its consortium partner, MEP Engineering, is a Spanish MEICA engineering firm that designs industrial piping drawings and 3D models daily.
Project info
SKIDGENT is a human-gated agentic AI system that takes SIAD’s compressor-skid piping work from design intent to simulation setup, using the inputs engineers already have: natural-language requirements, 2D piping drawings, 3D CAD assemblies and metadata.
Five agents run on a deterministic LangGraph orchestrator so that every step is logged and reproducible. A requirement-interpretation agent turns requests into a validated specification and asks targeted questions instead of inventing dimensions. A constraint-extraction agent combines an open-weight vision-language model, P&ID digitisation and geometric interrogation of the CAD assembly to build a typed constraint graph in which each constraint carries its source. A routing-decision agent drives established routing solvers with these constraints and treats skid size as an explicit optimisation goal. Geometry is built with CadQuery on OCCT and exported as STEP AP242, followed by automated meshing (Gmsh) and solving (FEniCSx/CalculiX). No number in any deliverable comes from a language model.
MEP Engineering engineers sign off at four expert gates, and their corrections become enforceable rules. Working targets, fixed at baselining, are a 40–60% cut in first-pass routing and setup effort and a 5–15% reduction in skid size.
Details
Routing compressor tubes inside a skid still depends on experienced engineers who read drawings in which many geometric and assembly constraints are only implied, then set up structural simulations by hand. Existing auto-routing tools only refine a route a person has already drawn, and no tool makes package size an explicit design objective, so conservative margins win and skids grow.
SKIDGENT, developed by Building Blocks with MEP Engineering, is an agentic AI system that extracts these partially explicit constraints from real drawings, CAD assemblies and metadata, takes the routing decisions itself with size as an objective, and passes the approved design into automated meshing, seed strategy and convergence setup. Language models read and propose; deterministic code calculates and verifies. Models are open-weight and self-hosted, so SIAD’s data stays within the approved environment. Practising piping engineers review the output at mandatory expert gates, and each correction they make is stored as a rule the system then enforces, preserving departmental know-how.
The project starts at TRL 2–3 and progresses through four gated milestones to TRL 5, with the engineering-effort ratio measured on unseen SIAD cases against Vendor Department baselines. Working targets are a 40–60% reduction in first-pass routing and setup effort, a 5–15% smaller skid, and constraint-extraction precision and recall of at least 0.90, with 100% compliance with expert sign-off. Smaller packages mean less steel, fittings and fabrication, and the consortium plans to bring the solution to compressor packagers and skid builders across the EU.
