BIMGUARD AI — Compliance Workflow

Phase 01 — Foundation

OpenBIM data ingestion

Weeks 1–3 · ~40 hrs ifcopenshell laspy open3d

Primary goal

Build a working OpenBIM data ingestion pipeline that reads MEP element data from any IFC 2×3 or IFC4 file and loads point cloud data from .las / .laz / .e57 formats — without touching any proprietary file format. Achieve >95% material normalisation accuracy across exports from Revit, ArchiCAD, Tekla, and Vectorworks.

Tools & stack

Python 3.11+ifcopenshell laspy (.las/.laz)open3d (.e57) pandasnumpy VS CodeIFC 2×3 / IFC4

Tasks

Export 3–5 IFC models from Revit, ArchiCAD, and Tekla to IFC2×3 and IFC4 format for cross-tool validationISO 16739-1
Write ifc_parser.py to extract element GUID, IFC type, material, system, floor/zone, XYZ centroid, and joint type via IfcRelAssociatesMaterialISO 19650
Build 22-material normalisation lookup table mapping all IFC label variants (e.g. "SS316L", "316 Stainless") to canonical BIMGUARD keys
Classify environment from IfcSpace zone names to T0–T5 wetting classes per EN ISO 15329 (e.g. "pool plant" → T5)EN ISO 15329
Write point_cloud_loader.py: .las/.laz via laspy, .e57 via open3d; KD-tree nearest-neighbour deviation query against IFC centroids
Build 25-element synthetic demo dataset with known risk outcomes including the critical SS316/pool-plant test case (0.00 galvanic / 0.89 crevice)
Validate parser output against manually reviewed IFC files; document false matches and unknown material flags

Deliverables

Moduleifc_parser.py

Modulepoint_cloud_loader.py

Datasynthetic_demo.py (25 elements)

Configmaterial_normalisation.json

Standards referenced

ISO 16739-1IFC open standard — primary model exchange format

ISO 19650BIM information management, property set naming

EN ISO 15329Wetting classes T0–T5, environment classification

ASPRS LAS 1.4Open point cloud format (.las/.laz)

ASTM E2807E57 point cloud format specification

Accuracy benchmark

Normalisation

>95%

Env. class

>90%

◈ OpenBIM argument: ifc_parser.py processes IFC from any authoring tool with identical logic. The compliance result for an ArchiCAD export and a Revit export of the same model should be the same — if it isn't, the difference is a finding about IFC data quality, not a limitation of BIMGUARD AI.

Phase 02 — Rule Engine

Corrosion compliance engines

Weeks 4–6 · ~35 hrs GC-001 CC-001 JSON rulesets

Primary goal

Build two independent compliance engines — GC-001 (galvanic corrosion) and CC-001 (crevice corrosion) — producing continuous risk scores 0–1 with four risk bands. Both engines run independently and in combination via combined_risk_assessment(). All 15 CLI validation scenarios must match known engineering outcomes.

Tools & stack

Pythonpandas numpyJSON (rulesets) NASA-STD-6012ASTM G48 EN ISO 15329CIRIA C692

Tasks

GC-001: implement 22-material galvanic potential table, voltage threshold logic (0.15 V harsh / 0.25 V normal / 0.50 V controlled), 5-band area ratio classification, PREN adequacy checksNASA-STD-6012
GC-001 scoring: weighted composite (0.50 × voltage risk) + (0.30 × area ratio) + (0.20 × environment multiplier)
CC-001: implement CCT table (SS304=−5°C through Titanium=120°C), 14-type joint library (JT-001 to JT-014), 7 environment severity classesASTM G48
CC-001 scoring: weighted composite (0.35 × geometry) + (0.40 × CCT adequacy) + (0.25 × environment severity)CIRIA C692
Version-control both JSON rulesets: galvanic_corrosion_ruleset.json and crevice_corrosion_ruleset.json at v1.0.0
Validate critical finding: SS316 flanges in pool plant room must score 0.00 galvanic / 0.89 crevice — proving dual-engine necessity

Risk band thresholds

Low

< 0.35

Medium

0.35–0.55

High

0.55–0.80

Critical

> 0.80

Deliverables

Enginebimguard_corrosion_engine.py

Enginebimguard_crevice_engine.py

Rulesetgalvanic_corrosion_ruleset.json

Rulesetcrevice_corrosion_ruleset.json

Standards referenced

NASA-STD-6012Galvanic voltage thresholds by environment class

ASTM G48 Method BCritical Crevice Corrosion Temperature (CCT) by grade

EN ISO 15329:2007Crevice testing, wetting class severity classification

CIRIA C692Stainless steel in construction, joint geometry guidance

IMOA Design ManualPREN formula and stainless steel grade selection

◈ Key innovation — dual mechanism detection: SS316 flanges in a pool plant room score 0.00 galvanic (no risk) but 0.89 crevice (Critical). A single-mechanism checker would miss this failure mode entirely. Two engines are required.

Phase 03 — AI Layer

LLM rule translation & AI interface

Weeks 7–9 · ~30 hrs Claude API JSON/SHACL Halo volumes

Primary goal

Use Claude API as a rule translation layer that converts building codes and engineering standards from natural language into machine-readable JSON/SHACL compliance rules. All generated rules pass human review before deployment. Build Halo clearance volumes for spatial proximity compliance. The LLM never makes risk decisions — White Box Architecture is preserved.

Tools & stack

PythonAnthropic Claude API JSON schemaSHACL n8n (optional RAG)ISO 19650 PSets Building Safety Act 2022

Tasks

Extract 10–15 applicable code clauses from ISO 19650, BS 8539, CIBSE Guide G, and Building Safety Act 2022 as source text
Build structured prompt template that converts code text to BIMGUARD rule JSON with: rule_id, conditions[], thresholds{}, standard_ref, risk_weight, human_reviewed fieldsWhite Box
Send clause text to Claude API, receive structured JSON, compare against manually coded rules — document divergence rate and root causes
Implement human review gate: all rules start as human_reviewed=false; a named reviewer must approve before promotion to production rulesetBSA 2022
Build Halo clearance volume geometry: AABB expansion around IFC element centroids by minimum clearance per material pair (50–100 mm depending on pair)
Build NL explanation generator: Claude produces a 2–3 sentence BCF violation description per element for the markup.Description field

Deliverables

Modulellm_rule_translator.py

Modulehalo_volume.py

Logrule_review_log.json

TemplatesNL explanation prompts

Standards referenced

ISO 19650BIM information management, Golden Thread provenance

Building Safety Act 2022Audit trail requirement for higher-risk buildings

BS 8539Fixings in construction, bi-metallic assembly rules

CIBSE Guide GMEP crevice corrosion guidance, pipe clearances

◈ Halo clearance volumes — key innovation: BIMGUARD AI generates spatial bounding volumes around each IFC element to check minimum clearance between dissimilar-metal assemblies. This enables spatial compliance checking that goes beyond point-to-point material pair analysis.

Phase 04 — Full System

Integration & BCF output

Weeks 10–13 · ~45 hrs Streamlit BCF 2.1 plotly

Primary goal

Build the complete 6-page Streamlit web application integrating all modules, and generate BCF 2.1 compliant issue exports with viewpoints, GUID linkages, engineer assignments, and schedule/cost impact data. Achieve <30 s processing time for a 100-element model. Run on the 25-element demo: target 9 Critical, 7 High, 9 Medium.

Tools & stack

Streamlitplotly pandasBCF 2.1 (buildingSMART) ifcopenshellzipfile ReportLabpython-docx

6 application pages

Data ingestion: upload IFC file OR load 25-element demo; upload point cloud OR auto-generate synthetic; single button to run compliance check
Model overview: all extracted MEP elements, filterable table, material inventory bar chart
Corrosion compliance: risk distribution pie chart, environment breakdown, 3D spatial risk map (plotly), full results table, CSV export
Point cloud comparison: IFC element centroids overlaid on point cloud in 3D coordinate space; deviation colouring
BCF issue manager: issues by band, engineer assignment chart, expandable issue cards, BCF 2.1 ZIP download
Schedule & cost impact: Gantt chart (baseline vs delayed), cost by band, delay by mechanism, issue-level breakdown

Demo run outcomes (25 elements)

Critical

9 issues

High

7 issues

Medium

9 issues

Cost impact

£170,600

Delay

162 days

Deliverables

Appapp.py (Streamlit, 6 pages)

Modulecompliance_runner.py

Modulebcf_generator.py

Moduleschedule_impact.py

OutputBCF 2.1 ZIP archive

OutputCSV asset register

Standards referenced

BCF 2.1buildingSMART issue tracking — viewpoints, GUID, history

ISO 16739-1IFC Pset write-back: Pset_CorrosionRisk, Pset_CreviceCorrosionRisk

ISO 19650Asset register format, information delivery specification

◈ BCF 2.1 vendor neutrality: Every issue exports with a buildingSMART-compliant viewpoint and GUID. The ZIP is importable into BIMcollab, Solibri, Navisworks, Tekla BIMsight, and any other BCF-capable platform — no proprietary viewer required.

Phase 05 — Extension

Evaluation & extension modules

Weeks 14–16 · ~35 hrs MC-001 MIC scikit-learn Fire module

Primary goal

Validate BIMGUARD AI against real IFC models with known corrosion failure history (precision ≥0.85, recall ≥0.80, F1 ≥0.82). Build MC-001 — Microbially Influenced Corrosion engine — as the third compliance module. Begin fire compartmentation module. Compile the master thesis document.

Tools & stack

Pythonscikit-learn matplotlib / seabornpandas PyTorch Geometric (GNN opt.)NetworkX python-docxpptxgenjs

Tasks

Test against 5+ real-world IFC models; calculate precision, recall, F1-score, confusion matrix per mechanism (GC-001 vs CC-001 separately)
Benchmark: BIMGUARD AI vs manual review vs Solibri — record time-to-detect, false positive rate, and missed issues
MC-001 MIC engine: dead-leg detection, stagnant water zone identification, under-insulation temperature modelling, SRB/IOB risk scoring for DN15+ pipework
Fire compartmentation module: IfcFireSuppressionTerminal checking, penetration seal compliance, cavity barrier clearance verification
Optional GNN: convert IFC spatial graph to NetworkX (rooms = nodes, adjacencies = edges), train on material proximity pairs to catch edge cases
Compile master thesis: merge Ch1–11 .docx files, unified formatting, abstract, executive summary, bibliography, table of contents

Evaluation targets

Precision

≥ 0.85

Recall

≥ 0.80

F1-score

≥ 0.82

Deliverables

Enginemc001_mic_engine.py

Modulefire_compartmentation.py

Reportevaluation_report.py

Thesismaster_thesis.docx

DeckFMP_presentation.pptx

OptionalGNN spatial engine

MC-001 scope — MIC mechanism

Dead-leg detectionFlow velocity <0.1 m/s in warm pipework (>20°C) — primary SRB risk

Under-insulationTrapped moisture beneath insulation on carbon steel — IOB and SRB

Stagnant zonesDN15+ pipe branches not on active circuits — Legionella crossover risk

◈ Three-mechanism coverage: GC-001 + CC-001 + MC-001 covers galvanic, crevice, and microbial failure modes — the three most common MEP corrosion mechanisms in UK commercial buildings. No existing OpenBIM tool covers all three in a single auditable pipeline.