How Broscorp Launched CityCode: An AI-Powered Regulations Expert for Architects Covering 20 Major French Cities

AI-powered regulatory compliance for architectswas the core problem Broscorp set out to solve when building CityCode. Broscorp developed an internal product to tame regulatory complexity, which later evolved into CityCode – a publicly launched AI regulations expertthat delivers instant, verifiable local planning constraints at the earliest stages of architectural design.

Results at a Glance

• Regulatory coverage across 20 major French cities
• 5,000+ urban planning and zoning documents indexed and structured
• >95% semantic answer accuracy across validated regulatory scenarios
• Automated QA and regression testing on every AI model update
• Successful transition from an internal AI knowledge system to a public enterprise-grade platform

Part I. Overview of the Project

Product Background

Broscorp builds complex digital products and internal platforms designed to function as an operating system for organizations – systems that help teams work with fragmented, high-stakes information in environments where errors are costly. During internal research, the team identified a persistent inefficiency in architectural workflows. Architects routinely spend significant time searching through local planning regulations, zoning codes, and municipal restrictions during the Preliminary Outline Design (PRE) phase. These constraints vary not only by city but often by zone or even specific location, while early design decisions critically depend on them. So, the challenge was not access to information. Most regulatory documents are publicly available. The real problem was interpretation, speed, and reliability. Architects needed precise answers to concrete regulatory questions, not document search results.

The Challenge

Architectural regulations are inherently complex to work with. They are highly localized, written in dense legal language, and distributed across large, unstructured documents. Discovering a regulatory conflict late in the design process often leads to costly redesigns and delays. Existing tools failed to address this reality. Document repositories and PDF aggregators improved access, but not understanding. Generic AI tools could generate plausible responses, yet lacked factual guarantees, traceability, and consistency – all essential for professional use. Traditional search interfaces also proved insufficient, as architects think in terms of constraints and scenarios rather than keywords.

Broscorp’s objective was to move beyond search and build an AI-powered architectural compliance software that delivers precise, verifiable answers aligned with real design workflows.

Project Goals

The mandate was defined around three clear objectives:

• Transform fragmented local regulations into structured, queryable knowledge;
• Deliver instant, city-specific constraints during early design stages;
• Achieve a level of accuracy, traceability, and consistency suitable for real architectural decision-making.

Part II. Technology Solution: Building CityCode as a Knowledge System, Not a Chatbot

A Bottom-Up Knowledge Architecture

Broscorp engineered CityCode using a bottom-up approach to AI knowledge management. Instead of treating regulations as static documents, the system models planning rules as structured, interpretable knowledge. Local planning documents are ingested and organized around zones, parameters, and regulatory constraints. A Retrieval-Augmented Generation (RAG) pipeline ensures that every response is grounded in authoritative source material. The language model does not generalize beyond retrieved content; it interprets specific regulatory sections and produces answers tied directly to official documents. This approach makes CityCode fundamentally different from generic AI chatbots and critical for reliable architectural regulatory compliance.

Quality Assurance and Validation Pipeline

Professional regulatory work requires more than plausible answers. CityCode, as an architectural intelligence platform, was built with a rigorous validation framework that treats correctness as a first-class requirement.

Proprietary Validation Dataset

Broscorp assembled a proprietary validation dataset comprising thousands of real regulatory questions drawn directly from official planning documents across 20 French cities. Each data point was manually verified and linked to its original regulatory source, establishing a transparent and auditable definition of correctness.

This dataset defines the reference baseline against which the CityCode system is continuously evaluated.

Automated Testing and Scoring

The CityCode system is tested nightly and on every new software version deployment. Each run evaluates not only whether an answer is semantically correct, but also whether it is appropriately supported by cited regulatory sources.

This automated QA pipeline prevents regressions, enforces consistency, and ensures the system remains reliable as regulatory coverage and model capabilities expand.

Continuous Improvement Loop

Any validation case that fails to meet defined accuracy or citation thresholds is automatically flagged. These cases are reviewed by the knowledge engineering team to identify weaknesses in retrieval logic, chunking strategy, or interpretation rules. Improvements are fed back into the system, creating a closed feedback loop that continuously strengthens factual accuracy and interpretability while preserving stability.

System Architecture Overview

CityCode is composed of four tightly integrated layers:

• A regulatory knowledge base containing structured local planning documents
• A retrieval and reasoning layer that interprets user intent and selects relevant context
• An LLM orchestration layer that generates constrained, fact-based answers
• A user interface designed for fast, precise regulatory queries and clear highlights of the source document

The resulting regulatory AI knowledge system is optimized for precision, traceability, and professional trust, rather than generic conversational output.

Part III. Results & Impact

CityCode successfully transitioned from an internal Broscorp system into a publicly available AI assistant for building rules and design limits, now accessible at CityCode. Architects can now validate local planning constraints in seconds rather than hours.

Internal testing across hundreds of real-world scenarios demonstrated:

• >95% semantic answer accuracy
• F1 score of 92% for source retrieval

These results confirm CityCode’s ability to deliver regulatory compliance AI for architects suitable for early-stage decision-making. The platform validated a broader concept: enterprise knowledge management powered by verifiable AI. The same architecture can be applied to other regulated domains where professionals rely on complex, evolving rule sets. For Broscorp, CityCode became both a market-ready product and a reusable foundation for future AI knowledge systems.

Building Verifiable AI for Real-World Knowledge Work

Managing knowledge today is not about storing documents. It is about enabling professionals to ask precise questions and receive answers they can trust. CityCode demonstrates how domain-aware retrieval, rigorous validation, and continuous quality assurance can turn generative AI into a reliable operational system rather than an experimental tool. If your organization struggles with regulatory complexity or fragmented knowledge, Broscorp designs and builds custom AI knowledge systems that scale with real-world complexity and professional demands.