A Design-Industry AI Platform Delivers Zero-Hallucination Specifications to Government and Enterprise Clients with Claude on Amazon Bedrock
At a glance
Enterprise clients in regulated industries need AI that produces accurate, citable answers. Standard RAG systems reduce hallucination without eliminating it, and a single fabricated reference carries real legal or compliance consequences. When a design-industry AI platform built a conversational AI layer for its enterprise clients, every pilot client rejected standard RAG on those grounds.
NMD built a production platform on AWS Bedrock pairing Claude with automated reasoning guardrails that validate every response against defined policies before delivery. A federal agency and a national enterprise manufacturer moved from explicitly rejecting AI to active production deployment.
The platform now runs production agent orchestration, multimodal search across text and image sources, Cognito-secured multi-tenant access, and an agentic retrieval loop for professional-grade accuracy. A third enterprise client is engaged for image-based product identification.
Industry
Use Case
Conversational AI / NLP; Intelligent Document Processing; AI Platform Migration
Solution implemented
- Claude on Amazon Bedrock for conversational AI with automated reasoning guardrails providing post-response validation against client-defined specification policies
- Amazon Bedrock Data Automation (BDA) for structured extraction of complex technical PDFs including diagrams, installation drawings, and multi-column specification tables
- Amazon OpenSearch Service for vector indexing of text chunks, PDF page embeddings, and product image embeddings
- YOLOv8 object detection in containerized ECS with LLM-based verification layer, reducing false positives on ambiguous product identification
- Amazon Bedrock AgentCore for production agent orchestration, natural language to SQL, graphRAG query routing, and response streaming
- Amazon Cognito for multi-tenant SSO and JWT-authenticated API access; DynamoDB for session memory; CloudWatch for observability; Terraform for infrastructure as code
The value equation
- Delivered zero-hallucination AI for specification queries in commercial construction and federal government contexts, enabling enterprise pilot conversion
- Unlocked production contracts with a federal government agency and a national commercial building products manufacturer
- Built multimodal image search that identifies physical products from user-uploaded photos against a manufacturer catalog
- Deployed a multi-tenant, Cognito-secured AI platform scalable across the client's full enterprise base
- Established AgentCore-powered agentic retrieval supporting natural language to SQL, graphRAG, and unified PDF and image search
Company Snapshot
The client provides an end-to-end AI and 3D visualization platform for the architecture, furniture, and interior design industries, enabling brands to power virtual showrooms, product configurators, and AI-driven knowledge and search experiences for their enterprise clients.
Location
Cambridge, Massachusetts, US
Customer Situation
Building products manufacturers serving commercial construction rely on technical professionals who answer detailed specification questions during live client engagements. For one enterprise pilot client, a wrong dimension or material specification on a large project loses the contract. That manufacturer tested a standard LLM internally and found roughly 90% accuracy. It was not enough. A federal agency piloting the same capability reached the same conclusion differently: in regulated procurement, a hallucinated specification is a compliance event, not an accuracy metric. Both clients were willing to deploy AI, but only at zero hallucination on policy-bound queries. Across commercial construction and regulated procurement, residual RAG error is not a tradeoff. It is a blocker.
NMD Solution
NMD reviewed and found the root cause: a single-pass retrieval architecture produced fluent, plausible responses but could not meet the zero-hallucination threshold required by enterprise commercial and government clients. The solution required pairing Claude on Bedrock with automated reasoning guardrails that validate LLM responses against client-defined specification policies before delivery to the user, Amazon Bedrock Data Automation for structured extraction of complex technical PDFs containing diagrams and installation drawings, and a multimodal dense retrieval index supporting text queries, image search, and natural language to SQL against product catalogs and specification libraries.
What We Delivered
Enterprise clients now deploy a production AI platform for specification queries and product identification. Bedrock Data Automation ingests technical PDFs, extracting diagrams and tables into OpenSearch vector chunks. Claude retrieves relevant content, generates a response, and automated reasoning guardrails validate it against defined policies before delivery, with page-level citations included. For product identification, YOLOv8 detects objects in user-uploaded photos, embeddings retrieve catalog matches via OpenSearch, and an LLM verification step reduces false positives. Cognito multi-tenant auth, AgentCore query routing, and CloudWatch observability complete the production stack.
Ready to Transform Your AI Software Business?
See how New Math Data can transform your AI software business with AWS-powered innovation.