Production,
not prototype.

Problem

A B2B scientific supplies distributor with 34,000+ products and a highly regulated catalogue (FUO/CLIA access controls) needed natural language search. Keyword search was failing: researchers couldn't find what they needed, and the system had no understanding of scientific terminology or product hierarchy.

Approach

A two-phase NLP and RAG pipeline. Phase 1 extracts semantic signals (category, filters, scientific intent) via RAG. Phase 2 applies them as intelligent boosts — not hard category locks — preventing hallucinated classification. 35+ synonym groups and 50+ product facets tuned for B2B scientific domain expertise.

Technical delivery

Monorepo architecture. Full-stack production build with semantic caching, real-time monitoring, and vector search. FUO/CLIA access control enforced throughout the search pipeline. CI/CD from day one.

Problem

A division of one of the world's largest industrial conglomerates manufactures precision sealing components for aerospace jet engines, nuclear reactor vessels, space launch systems, and deep-sea oil platforms — environments where seal failure is catastrophic. Their product catalogue (spring-energized seals rated to 10,000 PSI, metal seals to 79,770 PSI / 1,850°F, rotary lip seals, Rulon® fluoropolymers) previously required direct engineer involvement from the very first customer touchpoint.

Solution

An 8-agent AI system deployed on their website. A router classifies seal type from application description (temp range, pressure, motion type, leakage rate) and routes to specialist agents for Spring Energized Seals, Rotary Lip Seals, Metal Seals, a Seal Type Finder for unclassified enquiries, and a Sentinel agent that detects and bans prompt injection and social engineering. Vector search with neural reranking over embedded technical PDFs. RFQ capture routed to applications engineers.

Technical detail

The router is optimised for classification precision, not generation — no hallucinated responses possible at the routing layer. Hard "no hallucination" policy across all agents: must escalate to engineers rather than speculate on safety-critical specifications. UTM attribution tracking baked in. Separate EU/US data pipelines for GDPR compliance. A dedicated security sentinel persistently blocks prompt injection and social engineering attacks.

Origin

Started as a RAG experiment: a single AI retrieval chain over a Yamato oven error codes PDF, stored in a vector index. Lab managers had been calling support to look up error codes that lived buried in dense product manuals. The prototype worked. It evolved into a full 9-agent production system deployed on the manufacturer's US e-commerce site.

Solution

A 9-agent AI system with specialist agents for 6 product lines: laboratory ovens (18 model prefix families), water purifiers (Type I/II/III), autoclaves, rotary evaporators, muffle furnaces, and incubators. AI-powered knowledge search with inline citations rendered as clickable document links and PDF downloads. Each recommendation proactively followed by an upgrade suggestion. RFQ capture fires to human application sales experts.

Technical detail

Router trained on model number prefixes (DX/DKN/DNE = ovens; WA/WB/WC = water purifiers) for fast zero-cost classification. Edge cases from real customer enquiries encoded in training data: 427°C+ oven requests, RS-485 serial communication for automated test environments, below-room-temperature incubation, UV oxidation water purifiers for DNA/PCR applications. Multilingual (Korean, Spanish, French). One-question-at-a-time conversation design throughout all agents.