AI Engineering · India
JSW BuildAssist - An AI-Powered Home Construction Planning and Estimation Wizard
A web wizard that turns a homeowner’s floor plan into a material list, a location-specific cost estimate and a list of nearby authorized dealers, with an AI chatbot available throughout.
Discuss Your AI Project- Floor Plan Reading
- City-Specific Cost Estimates
- Dealer Suggestions
- AI Chat Assistant
About the client
JSW is one of India’s leading manufacturers and distributors of building materials, with an extensive network of authorized dealers across multiple cities. Its independent home builder division serves homeowners planning to construct their own residences, a segment that is large, underserved, and highly dependent on manual, fragmented processes for planning and procurement.
- Industry
- Building Materials / Construction Technology
- App type
- AI-powered home construction planning and estimation wizard (web)
- Practice
- AI Engineering (Build)
- Services
- AI Product Engineering, Agents & Agentic Workflows, AI Quality Engineering & Evals, AI Infrastructure & LLMOps, AI Platform Partnerships & Professional Services, Software & Application Development, AI Experience Design (UI/UX), AI Strategy, Governance & Compliance Advisory
Tech stack
- Orchestration and agents: LangGraph (stateful agent orchestration), LangChain (ChatGoogleGenerativeAI, SystemMessage, HumanMessage, multimodal content assembly)
- AI layer: Google Gemini (via langchain-google-genai), for multimodal reasoning across the text layer and visual renders
- Observability: LangSmith (LLM call tracing and prompt iteration)
- Backend: FastAPI + Uvicorn (async REST API across auth, upload, estimation, dealers, projects and chat), Python (agent logic, PDF parsing with PyMuPDF, image processing with Pillow)
- Database and migrations: PostgreSQL (LangGraph checkpoints plus application tables), SQLAlchemy Core + Alembic (schema management and versioned migrations), TypedDict (canonical row contracts in code)
- Frontend: Next.js 15 (App Router), TailwindCSS v4, Framer Motion (animated step transitions)
- Hosting: GCP Cloud Run (asia-south1, Mumbai), Cloud Storage
Engagement summary
We built BuildAssist, an end-to-end web wizard where a homeowner uploads an architectural floor plan and goes through five steps: upload file, project insights, material estimations, cost estimator and dealer suggestions.
The AI agent extracts the project specifications from the plan, and users can review and correct the extracted data before going on. Material estimation, session management and dealer lookup are handled by deterministic systems.
An AI chatbot is available throughout the wizard. It answers questions about the user’s own uploaded plan or saved project, and general questions about JSW’s product lineups on the BuildSure platform. Users have accounts, with session persistence and project saving.
Challenges faced by the Customer
Fragmented and manual planning process
Homeowners had no unified tool to go from a floor plan to a material list and cost estimate. Each step (reading the design, computing quantities, researching costs, finding dealers) was done manually, often with multiple consultants, leading to inaccuracies and delays.
Architectural drawings are hard for machines to read
Floor plans come in common formats: scanned PDFs and images (JPG, PNG, WEBP). They contain information at multiple levels, including pre-computed area schedule tables, handwritten or printed text annotations, and visual geometry. No single machine extraction strategy works for all of them.
Subjective interpretation of design documents
A drawing may show dimensions in text, on the visual render, or in an area summary table. These details may be missing or may not match, so the system needs to compare them and resolve any differences rather than rely on a single source. The system needs to reconcile multiple sources of information instead of trusting a single pass.
No structured data from unstructured inputs
Homeowners upload a PDF or image. The output needed is precise structured data: total area in sq ft, number of floors, configuration, structural type, parking and utility areas, and per-floor room counts.
Material estimation needs local context across a diverse, pan-India market
Material prices, labour rates and cost distributions vary significantly by geography and by the tier of the city or town. A one-size-fits-all estimate isn’t useful, so the system needed a location-aware formula engine that computes accurate quantities and costs for every state and union territory, across tier-1, tier-2 and tier-3 markets.
Dealer discovery is disconnected from the planning journey
After estimating materials, homeowners had no easy way to find authorized dealers nearby. The existing process required separate enquiries outside the platform.
Results
End-to-end planning time reduced from days of manual consultation to minutes of self-serve AI-wizard use
Dealer discovery built directly into the post-estimation step, removing the need for a separate search
Session persistence and project saving let users pick up where they left off, across devices
Two AI chat surfaces in one experience: project-specific guidance and general BuildSure website and JSW product Q&A, both grounded in the same knowledge base
Pan-India, tier-aware cost estimation: every state and union territory is mapped to representative tier-1, tier-2 and tier-3 cities with estimated rate data, and users outside that list are classified by state or union territory and tier (never by geographic proximity) and matched accordingly
BuildAssist is designed to convert a passive research visit into a qualified, attributed lead. Because every material estimate spans JSW’s portfolio in one pass, the wizard naturally cross-sells product lines a homeowner may not have been actively researching. The dealer-suggestion step at the end of the estimate shortens the path from “I have a plan” to “I know who to buy from”, which supports dealer network activation and engagement. The always-available chatbot keeps visitors engaged with JSW’s product knowledge at the point where they make buying decisions.
How it works
- Cost estimation, dealer lookup, session management and project persistence are entirely deterministic: fast, cheap and auditable. The LLM handles only the parts that need genuine reasoning, like interpreting drawings and reconciling ambiguous input. The calculator doesn’t call the LLM; the chatbots do.
- Database row shapes are defined once as TypedDicts and mirrored through SQLAlchemy and Alembic migrations, so there is no implicit schema drift between the agent, the API and the database.
- Text and image content from the same uploaded document are combined into a single request to Gemini, instead of being processed as separate passes.
Core solution
Workflow decomposition
We broke the homebuilder journey into atomic steps and identified which segments are deterministic (formula calculations, database queries, auth) and which are agentic (document understanding, context extraction). Keeping LLM usage targeted and cost-efficient.
Design analysis agent
The floor-plan analysis pipeline is a LangGraph-orchestrated agent. When a user uploads a floor plan, the file is first validated for extension, size and file signature, then parsed for both its text layer and its visual page renders. Gemini 2.5 Pro performs multimodal extraction of the project metadata, and the result is validated and persisted as a design submission before it is handed to the deterministic formula engine. Uploaded design files are stored in a dedicated Cloud Storage bucket. Row contracts throughout the system are expressed as TypedDicts, keeping the schema consistent from extraction through to the database.
Multi-source document intelligence
Each uploaded document is processed for both its text layer and its visual page renders (PyMuPDF for PDFs, Pillow for images), which gives Gemini multiple complementary sources of information in a single extraction call, instead of relying on OCR or vision alone.
Iterative prompt engineering with LangSmith
LangSmith tracing let us inspect extraction attempts and the model’s output, which we used to refine the system prompt step by step instead of running blind.
The formula engine
A location-aware Python module computes material quantities (steel in MT, cement in bags, paint in litres), location-specific labour costs by phase, phase costs distributed across construction phases as a fixed percentage of total material cost, and quality tiers (standard, premium and luxury multipliers). It looks up tier-based city rate tables in the database. This is entirely deterministic, with no LLM involved anywhere in the path, which keeps the estimation endpoint fast, cheap and auditable.
Pan-India, tier-based location coverage
Every state and union territory is mapped to a representative tier-1, tier-2 and tier-3 city, each backed by estimated material and labour rate data. Users pick their city from a curated dropdown, or type in a city that isn’t listed. In that case, the AI agent classifies it by state or union territory and by tier, then applies the rate data of the existing representative city that shares that same state and tier. This is a classification match, instead of a typical a distance-based lookup.
Dealer suggestions
After the estimate, the wizard suggests authorized dealers near the user’s pincode, through a deterministic lookup.
AI chatbot, available throughout the wizard
A LangGraph-backed chat agent answers two kinds of questions in one surface: project-specific queries about the user’s own uploaded plan or saved project, and general queries about JSW’s product lineups on the BuildSure platform. Both are grounded in the same product knowledge base and powered by Google Gemini (Gemini 2.5 Flash), with conversation state persisted as PostgreSQL checkpoints so sessions survive across wizard steps.
Service and deployment
BuildAssist runs on a dedicated FastAPI AI agent service, called over REST by the Next.js 15 wizard. The three capabilities (design analysis, material and cost estimation, and the chat assistants) are separated by whether they need an LLM, and they all read from and write to the same PostgreSQL schema, with LLM calls instrumented with LangSmith for tracing. The platform is deployed on GCP Cloud Run in the asia-south1 (Mumbai) region with per-IP rate limiting.
More Case Studies
AI Engineering · United StatesVTS Rise
VTS Rise · Real Estate, PropTechVTS Rise is a property operations and tenant experience platform, that digitizes the entire asset experience, for landlords. It enables asset and property managers to power their portfolio-wide operations and provides a premium experience to residents, building operators, and visitors.Read the Full Case Study
AI Engineering · IndiaWondrAgents - An AI-Powered Multi-Agent Platform for Influencer Marketing Operations
Opportune · Influencer Marketing / AdTechAn AI-powered platform that brings the entire influencer marketing lifecycle into one automated workflow, from creator discovery to campaign execution and evaluation.Read the Full Case Study
AI Engineering · New ZealandA LangGraph Platform for AI-Powered Location Intelligence
MarketingWe built five production AI workflows on a unified LangGraph + Gemini stack, improving speed with early-return results and URL-level caching. The system supports resumable conversations, enforces governance within AI nodes, uses optimized multi-layer storage, and ensures reliability through schema-validated outputs—all accelerated by reusing Reach’s existing AI framework.Read the Full Case Study
Let’s Explore How We Can Build Together
Whether you’re building a new product or modernizing an existing one, get insights from our product, engineering, and AI experts.