Case studies

Project Memory System

The deployed prototype of the methodology behind every Project Brain Build. Solo work, in daily production since 2025.

Python · Claude · Obsidian · SQLite FTS5 · cron — solo

The problem

AEC firms accumulate a decade of context — emails, drawings, contracts, meeting notes, regulatory documents — that no one can retrieve when it matters. By the time a partner needs “what did we agree with client X about Y last March?”, it’s faster to forget the question than find the answer. Before building project-memory systems for firms, I built and operate one on my own life.

What I built

• Ingestion: Gmail every 4 h, Calendar daily, WhatsApp full-history, ChatGPT/Claude exports, meeting transcripts, a local-file watcher — all custom Python, all on cron, all idempotent, all locally stored.
• Curation: 11 categorized vaults, ~305 MB curated context, source stubs tracing every note to its original file, append-only decision log, SQLite FTS5 retrieval index, quarantine folders — nothing deleted, only moved.
• Multi-agent layer: 12 specialized AI skills — curator, researcher, daily, blind-spots, architect, vault-cleaner, orchestrator and more.
• Automation: morning brief 7 a.m., daily note 10 p.m., weekly review Sunday.
• Governance: local-first, zero-data-retention DPA, aligned with EU AI Act art. 4 + art. 50.

7,571 lines of Python · 17 production scripts · 12 AI skills · 11 vaults. Last architecture audit examined 8,775 notes and produced a prioritized fix-it queue.

Why it matters for your project

The transfer is mechanical: email → project-management tool, chat → team channel, files → BIM/document archive. The runtime risk is low because the system already runs in production daily — the work is connector substitution and prompt re-targeting, measured in weeks, not months.

Green Score — Embodied Carbon from BIM

A 2-agent CrewAI system that turns an IFC model and project documents into a defensible embodied-carbon report in minutes. Solo R&D.

CrewAI · Claude Sonnet · IfcOpenShell · ICE Database v3 — solo

The problem

Teams need the carbon footprint of a design before construction — but the numbers live in three disconnected places: the BIM model, the spec book, and the quantity takeoff. Reconciling them by hand is a junior-engineer week, so most projects skip it.

What I built

Agent 1 — BIM Quantity Analyst reads the IFC, extracts net volumes per element, cross-references the memorial descritivo. Agent 2 — Embodied Carbon Specialist applies ICE v3 coefficients, computes total kg CO₂e and a Green Score, identifies hotspots, and writes a DOCX report with costed mitigation actions. Sequential crew — the reasoning chain stays auditable.

Outcome — real 186 m² residence, Recife

• Green Score 42/100 · total embodied carbon 102,410 kg CO₂e (549 kg/m²)
• Hotspots: rebar steel 39.9% · ceramic masonry 17.6% · concrete structure 14.5%
• Five mitigation actions, each with projected impact, cost, and ROI

Combined effect: −17.7% embodied carbon, score 42 → 58, for ~2.2% of project cost. Every number traces to a model volume and a published coefficient.

Bret & Rachel — A Residential BIM Rebuild

A two-dimensional drawing set rebuilt as a full Revit BIM model and visualised in photoreal renders.

AutoCAD DWG → Autodesk Revit → photoreal rendering — solo

The problem

The clients had a complete 2D CAD set — adequate for permits, illegible to a non-technical eye. A DWG describes geometry; it doesn’t describe volume, light, or how the porch reads from the street at four in the afternoon. They needed to see the house before it was built.

What I built

A complete, fully parametric Revit model — layered walls, composite roof, families, porch, detached garage, site from aerial imagery — plus a set of photoreal daylight exterior renders. A change in the brief propagates through every plan, elevation, and render automatically.

Why it matters

A drawing set is a contract. A BIM model is a conversation. Material swaps cost minutes in Revit and zero on site; proportion problems get caught on screen, not after framing. For renovation or visualisation work on legacy-CAD properties, the same pipeline transfers directly.

Multi-Agent BIM Acoustic Coordination

A 7-agent system verifying acoustic compliance on BIM models against Portuguese regulation (RRAE / DL 96/2008). Zigurat Institute, in active development.

IFC · BCF · ifcopenshell · RAG · JSONL decision memory — team of five

The problem

Acoustic compliance data lives in three disconnected places — the IFC model, the acoustic report PDF, and the memorial descritivo. Coordination today is manual cross-reference at the end of design, when changes are expensive. No openBIM-native tool closes the loop.

What we’re building

Seven agents — orchestrator, IFC reader, report parser, deterministic verifier, decision logger, history RAG, report writer. Three of seven are deterministic by design: separation by responsibility, not technology. Test case: a real Portuguese residential development. Deliverables: IFC + BCF + DOCX, with versioned decision history that BCF alone doesn’t preserve.

My contribution

• RAG research and the IFC-vs-Revit technology consolidation
• “Pipeline TFM com Claude” — the execution-architecture proposal: two-repo separation, per-agent LLM/deterministic mapping, 12-week roadmap, decision-memory schema
• Mapping a 221-skill AEC AI landscape onto the 7-agent architecture

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

Why it matters for your project

The same pattern transfers to any EU-regulated compliance check — thermal, energy, accessibility: extract from IFC, parse requirements, compare deterministically, log decisions, generate auditable deliverables.

Multi-Agent ABNT Compliance Audit

A 5-agent CrewAI crew that audits a residential project’s memorial descritivo, analytical budget, and IFC model against three ABNT standards — then rewrites each document with corrections traced to specific norm clauses.

CrewAI · Claude Sonnet + Haiku · IfcOpenShell · python-docx · openpyxl — solo

The problem

On every Brazilian residential project, three documents have to agree — the memorial, the budget, and the IFC. They never do. The memorial cites NBR 6118:2014 while the structure assumes 2023; the budget bills 850 m² of masonry but the IFC walls add to 920 m². Each gap is a future RFI or an aditivo contratual. Manual reconciliation against NBR 9050, NBR 15575, and NBR 6118 is a senior-engineer week — so it gets skipped.

What I built

A crew mirroring a real coordination office: a project manager who delegates and validates, a technical auditor who issues the conformity report, and three discipline specialists who each own a deliverable. Mixed-model for cost control. Output: a formal compliance report with clause-level non-conformities, plus the three input documents rewritten with corrections applied.

EU AI Act — Governance for Multi-Agent AEC Systems

A formal regulatory and ethical evaluation of CrewAI-based multi-agent systems in AEC/BIM workflows, with a deployable 7-directive governance protocol for Portuguese and EU offices.

Regulation 2024/1689 · RGPD · IEEE 7000 — solo

The problem

The EU AI Act enters full application on 2 August 2026, with deadlines already triggered since February 2025. Penalty exposure for GPAI providers reaches €15M or 3% of global turnover — and deployers carry the framework downstream. Most AEC offices haven’t classified their current AI usage, let alone governed what they’re about to deploy.

What I built

A three-layer classification model separating the framework (CrewAI), the foundation model, and the deployed system — three distinct regulatory objects, with Portuguese professional liability layered on top. The same codebase can be limited-risk for one use case and high-risk for another; classification is dynamic. The deliverable is a 7-directive governance protocol an office can actually adopt, not an academic essay.

James & Charlotte — A Carvoeiro Villa, Permitted from the Plot Up

A UK couple close to buying a 2,400 m² rural plot in Lagoa, with a villa already sketched. A pre-acquisition PIP rewrote the brief before any deeds were signed.

PIP · PDM · REN · PMDFCI · ETAR · RJUE — worked example

The plot looked clean; nobody could answer whether the sketched villa was permittable on it. A Pedido de Informação Prévia filed with the Câmara Municipal de Lagoa returned three constraints the original team hadn’t accounted for — a 192 m² buildable-footprint cap, a 50-metre fire-defence buffer, and a mandatory ETAR (~€18,000).

Result in the worked example: plot renegotiated €490,000 → €452,000, reshaped bilingual brief, and an issued alvará de obras in 22 weeks against a 9–14 month Algarve baseline.

Daniel & Erin — When a Lisbon Refresh Became a Reconstrução

An American couple bought a small Campo de Ourique building intending a three-month interior refresh. The Câmara reclassified the works as reconstrução — triggering the full RJUE process.

RJUE · SCE · SCIE · RRAE · comunicação prévia → alvará — worked example

What looked like an interior refresh crossed the legal line into reconstrução, invalidating the original comunicação prévia. In the worked example, a bilingual project memory — every drawing, parecer, and decision indexed and queryable — takes the licence from rejected comunicação prévia to issued alvará in eight weeks, with the American owners able to follow every step in English.