Inada Manufacturing — Quote turnaround: 6 hours to 48 seconds

Inada Manufacturing produces high-precision massage chairs sold into hotel, medical, and retail markets globally. Each product line supports hundreds of configuration variables — motor specifications, upholstery options, localization requirements, logistics zones, and custom branding.

A quote for a custom configuration required a sales engineer to manually cross-reference three internal pricing spreadsheets, a logistics rate table, a custom options matrix, and exchange rate feeds. The process took six hours per inquiry. At peak inquiry volume, this created a 3–5 day backlog.

Inada needed a system that could ingest incoming specification documents — often PDFs or structured emails from distributors — and return a complete, accurate quote without human intervention.

The pipeline uses Gemini 2.5 Flash for document understanding — its vision capability handles non-standard PDF layouts, hand-annotated spec sheets, and mixed-language inputs. Claude 3.5 Sonnet handles the reasoning layer: applying Inada's pricing rules, resolving ambiguous option combinations, and flagging edge cases for human review rather than guessing.

Every decision in the pipeline is logged to MongoDB with the model's confidence scores and the input state — giving Inada a complete audit trail for every quote, whether or not a human reviewed it.

QuoteForge went live in Q3 2024. Quote turnaround dropped from an average of six hours to 48 seconds. Backlog cleared within two weeks of deployment. The sales team reports handling three times the inquiry volume with the same headcount.

Edge cases — configurations the model flags with low confidence — are routed to a human review queue. This accounts for approximately 4% of inquiries and takes 20–30 minutes (a reduction from 6 hours) because the model pre-fills everything it can.

The initial version used a single model for both extraction and reasoning. Splitting these into two discrete model calls — one fast/cheap for extraction, one capable for reasoning — reduced latency by 30% and cost per quote by 45%.

We'd also build the human review queue before the first deployment rather than retrofitting it. The initial assumption was that confidence scores alone would be sufficient signal — they're not. A structured review interface should be part of v1.