A 7-agent system that verified acoustic compliance on BIM models against Portuguese regulation (RRAE / DL 96/2008). Team master's thesis completed at Zigurat.

The Problem

Acoustic compliance in Portuguese residential construction is governed by the RRAE (Regulation of Acoustic Requirements for Buildings, DL 96/2008) — measurable requirements (Rw, D₂m,nT,w, L'nT,w) that walls, slabs, and facades must meet.

The data needed to verify compliance traditionally lived in three disconnected places:

The BIM model (IFC) — wall types, layer compositions, dimensions.

The acoustic project report (PDF) — required values per wall type.

The specifications document — material specifications and assumptions.

Coordination historically required manual cross-referencing, performed at the end of design when changes were expensive. We built an openBIM-native tool to close the loop: parsing the IFC, reading the requirements, running the comparison, logging the non-conformity as a BCF issue, and preserving the decision history across model revisions.

What We Built (Team)
We developed a 7-agent system that ingested an IFC model and an acoustic report, verified each wall against its requirement, logged structured decisions, and produced an IFC + BCF + DOCX deliverable.

Agent 00 — Coordinator: Responsible for orchestration and state management (LLM).

Agent 01 — Model Reader: Handled IFC parsing via ifcopenshell and extracted wall Psets (Deterministic).

Agent 02 — Report Reader: Converted PDFs into structured JSON requirements (LLM).

Agent 03 — Verifier: Executed the numerical comparison ensuring Rw_real ≥ Rw_required (Deterministic).

Agent 04 — Logger: Maintained an append-only JSONL decision log (Deterministic).

Agent 05 — History: Managed RAG over decision history for semantic recall (LLM).

Agent 06 — Report Generator: Produced the narrative report generation in DOCX format (LLM).

Three of the seven agents were deterministic by design. Our academic argument was that separation by responsibility, not by technology, is what makes the architecture genuinely multi-agent rather than just an "LLM with prompts."

Test case: Quinta das Acácias (project T2022-335 / 7510) — a real Portuguese residential development.

Standards: IFC, BCF, IDS — all openBIM. The shared decision memory preserved what standard BCF does not: versioned reasoning across model revisions.

Team: Bárbara Bernardo, Daniel Veludo, Luis Miguel Acedo, Naiara Forneck, Tatiana Pedrosa. Supervised by Rogério Lima.

My Contribution
I delivered three discrete pieces of individual work within the group:

RAG research and IFC-vs-Revit consolidation: Reviewed Lewis et al. (2020) and Chen (2025) RAG4CM to assess how retrieval-augmented generation applied to agents 02 and 05. Consolidated the case for utilizing IFC and ifcopenshell over the native Revit API.

Master's Thesis Pipeline with Claude — execution architecture proposal: Authored the strategy document for executing the entire project on Claude Code, subagents, and the Claude API using a local-first approach. This included designing a two-repository separation, the per-agent LLM-vs-deterministic mapping, a 12-week roadmap, a JSONL schema for the shared decision memory, scoping the initial prototype to one wall by one requirement end-to-end, and drafting a six-item risk register.

AEC AI skill landscape mapping: Mapped the Data-Driven Construction collection of 221 AI skills for AEC onto the 7-agent architecture. I identified adaptable skills (IFC parsing, PDF to structured data, BIM validation pipeline, RAG, document generation) and highlighted where the team's contribution remained original (decision history preservation, multi-version reasoning, and addressing the BCF gap).

Outcomes & Relevance
The project was successfully built, implemented, and defended.

For firms invested in openBIM and EU-regulated compliance work — acoustic, thermal, energy, accessibility — this architecture pattern is highly transferable. The system extracts from IFC, parses requirements from documents, runs deterministic comparisons, logs decisions with versioning, and generates auditable deliverables. The specific RRAE regulations used in our system are fully interchangeable with REH, RSECE, NBR, or any IDS-expressible requirement set.