Instant Clarity
Plain-English question and answers to complex zoning rules.
Client Overview
Client: Confidential (U.S. Based AI Startup)
Location: United State
Industry: Urban Planning, Civic Technology, GovTech
Project Background
In the United States, planning for housing development becomes a time-consuming process due to the complexity of zoning regulations and land use documents. These dense materials can create delays and confusion. The client aimed to create a user-friendly platform that allows users to ask natural-language questions about their property and receive clear, instant answers regarding applicable restrictions. The goal was to develop a cost-effective system to simplify the zoning interpretation process and make regulatory information more accessible to everyday users.
Technical Challenges
Unstructured Regulations : Zoning information is buried in dense, unstructured PDF documents written in legal language.
Fragmented Data Sources : Critical property data (e.g., zoning codes, dimensions) is spread across siloed systems like ZIMAS, geocoding databases, and public records.
Ambiguity in Legal Language : Interpretation of zoning codes requires contextual understanding and logical reasoning.
Token Cos : Using LLM to calculate regulations from user-provided parameters is too expensive, as it requires regulations to be understood in full.
Hallucinations : Data hallucinations occur where documents are written to reference the sections like “refer section 2.1” which if not considered it can give incorrect answers
Technical Implementation
Data Formatting and Normalization
Zoning documents in HTML, DOCX, and PDF were parsed, structured, and normalized. Metadata and references were captured to enhance searchability, filtering, and contextual relevance.
Retrieval-Augmented Generation with Caching
User queries matched via vector similarity; relevant chunks passed to LLM for responses. Summarized chunks cached as metadata to reduce payload and improve response efficiency.
Agentic Workflow Orchestration (LangGraph)
We implemented an Agentic Workflow that decides which tools to use based on user queries. It utilizes rules engine to compute user-given property information or uses the RAG and fetches textual information.
Rules Engine to Escape LLM Calls
Created Rules-Engine to avoid LLM calls to calculate the property regulations and restrictions using predefined logic and validation checks.
Business Benefits
Time Savings
Avoids hours of manual document sifting or city office visits.
Own Property Zoning Analysis
Users can input their property address or details to receive tailored zoning insights, helping them understand applicable regulations and development constraints with minimal effort.
Low-cost Implementation
Using the Rules-Engine prevented the need of RAG with LLM and thus impacted the cost on a large scale.
Key Innovation
Extracted and tagged location-specific metadata (section titles, paragraphs, tables, and zoning terms) during data ingestion to enable precise filtering and relevance in responses. Since regulations were common per the sector, we created Rules-Engine that can be called by Agent to reduce the LLM calls. Furthermore, to reduce the token cost, we created embeddings with the cached summaries. We wrapped the LLM prompts with hallucination-safe checks. Also, the data was tagged with metadata that lead to immediate and correct queries.
Architecture
Solution
Built a modular ingestion system to fetch, clean, and structure documents from government sites and planning departments, preserving source fidelity. Used embedding models to chunk zoning documents and store them in a vector database for fast semantic retrieval based on user queries. Integrated an LLM orchestrator to formulate responses that include references to the original document, enhancing trust and verifiability. Integrated regulations with Rules Engine to avoid unnecessary LLM calls for similar calculations.
Conclusion
Working on this project, we combined our technical expertise with emerging tools like LLMs and agentic workflows to solve the challenges of unstructured zoning data and legal ambiguity. Despite the fast-evolving nature of these technologies, our team adapted quickly, building a cost-efficient, reliable system using rules engines, smart caching, and thoughtful orchestration. Throughout, we stayed focused on delivering a smooth, trustworthy experience for end users.
They loved working with Coditude!
We act as a global catalyst for change, empowering our clients and partners to leverage technology, unlock growth, achieve sustainability, and bolster capabilities. See how our clients benefited from our partnerships.
Impact Stories: From Vision to Product
Case Study
Pioneering Progress
The client, a market leader in immigration and visa services with a vast network spanning 60 offices in 27 countries, sought to revamp their outdated digital infrastructure.
View Case Study
