Two Pillars of Manufacturing AI Deployment: Drawing Review + Quotation Loop

The real reason generic AI doesn't fit factories isn't that "the model isn't powerful enough" — it's that "factory data is scattered, rules are messy, and no one outside the team understands the sign-off logic." Nox-Lumen Mfg doesn't bet on AGI. Instead, it makes "machines can sign off, wrong orders get intercepted" the core loop for engineering procurement businesses.

Why a "Dual System" Rather Than Two Separate Tools

The chain from inquiry to production in engineering procurement naturally has two segments: drawing review → quotation. The inputs and outputs of these two segments are tightly coupled:

• The "structured window/door list" output from drawing review is exactly the input for quotation
• "Client preferences" for the same client are used by both segments
• The "red hard-stop" from the drawing engineer must be able to block the quotation system from accepting the order

If the two segments are separate tools, the interface is manually transferred, and wrong orders will still flow from drawing review to quotation, and from quotation to the factory. If both segments are modules in the same platform, they're automatically linked through file + state contracts, and wrong orders are intercepted at the source.

Drawing Review System: Free Engineers from "Flipping Through Drawings"

One-Line Description

Machine does the initial review; engineer does the final review. Frees drawing review engineers from "going through each drawing to verify dimensions/quantities/glass/hardware" — they only review the critical issues the machine flags. Client preferences and error patterns accumulate as reusable assets.

The system never replaces an engineer's sign-off. It never "acts on its own initiative" — it only asks questions; it never suggests values.

Core Capabilities (M1–M10)

Multi-Source Document Parsing

Key design: Every extracted field carries source + confidence. Low-confidence fields don't enter the main results directly — they require manual confirmation by sales / engineer. Never guess values on low-quality input.

Multi-View Reconciliation

Every reconciliation issue must list all source values + evidence pointers (which cell / which page, which region) + cropped image evidence.

Three-Layer Rule Library

Every rule is visible and auditable: trigger conditions, severity, violation message, source. Client/project-level rules entered once, applied forever.

Issue Classification and Hard Gates

Hard constraint: The system never allows a project with unresolved red issues to flow to quotation / production.

See Drawing Review Solution.

Quotation System: Turn 2,000 Historical Quotes into Searchable Assets

One-Line Description

Give sales reps a project-level composite $/m² recommendation (including add-ons), guaranteed not to lose money, aligned with similar historical projects. Transform 2,000 scattered historical quotes into a system-searchable knowledge asset.

The system never replaces sales rep judgment. It never builds factory BOM — factory ¥ is a black box input; the system doesn't replicate or intervene in factory internal cost calculations.

Project Tier Classifier

Projects are classified into 5 tiers by total square meters. Each tier learns independently to prevent large/small project strategies from contaminating each other:

Historical Memory Store

Recommended Price Engine

Recommended price = max(
    breakeven floor,
    historical neighbor $/m² × tier coefficient × client type coefficient × add-on coefficient
)

Every recommended number is traceable to:

• Factory rows (which BOM lines)
• Tier (XS/S/M/L/XL)
• Coefficient source (client type, add-ons, market adjustment)
• Historical neighbors (specific project IDs)

Breakeven Constraint

See Quotation Solution.

Dual-System Hard Contract Integration

See Dual-System Hard Contract Integration.

Shared Historical Memory Store

Both systems share one client-isolated memory store:

How to Get Started (Dual System Doesn't Need to Go Live Simultaneously)

Although dual-system integration works best, both systems are decoupled through file + state contracts and can go live independently:

Recommended Implementation Sequence

Week 1-2:   P0 (sample governance, shared)
Week 3-5:   Drawing Review P1 + Quotation P1 (parallel)
Week 6-8:   Drawing Review P2 + Quotation P2 (parallel)
Week 9:     Integration testing (Drawing Review → Quotation)
Week 10-11: Experience accumulation (shared memory store) + company-wide go-live

Total duration: 10–11 weeks — nearly the same as a single system (shared P0 + parallel implementation).

FAQ

Q: Our drawings are in German / French — can the system parse them? A: Yes. OCR + multilingual model handles them. Accuracy depends on drawing quality. We recommend running 20 samples as a baseline during P1.

Q: We already have an ERP / MRP system — can this integrate? A: Yes. The structured list from approved drawings can interface with ERP via API / file export (SAP / Odoo / UFIDA / Kingdee / custom builds all supported).

Q: Can it run fully on-premises, inside the factory network? A: Yes. L1 parsing can run fully locally; LLM uses the customer's intranet LLM endpoint. See Private Deployment.

Q: Who maintains the client preference library? A: Project managers maintain it in the system (there's a "Client Preference Management" module in the UI). Bulk import from Excel / email is also supported.

Full cases + contract design + memory store architecture at docs/solutions/manufacturing.