This post walks through the actual dual-system implementation of Nox-Lumen Mfg for a real engineering procurement client. Every factory has different processes — what follows is for reference only, not directly replicable. The actual solution follows: P0 pain-point diagnosis → P1 PoC → P2 customization → P3 managed operation.
Business Background
| Dimension | Scale |
|---|---|
| Business type | International window engineering procurement |
| Cycle | Average 2–4 weeks from client inquiry to production order |
| Historical projects | ~2,000 scattered Excel / PDF files |
| Client types | Mix of international developers + engineering contractors |
| Team | 6 sales reps + 4 drawing review engineers + 2 project managers |
Pain Point List
| Pain Point | Current State | Loss |
|---|---|---|
| Slow drawing review response | Engineers review when they're free; multiple back-and-forths | Lost orders |
| Drawing engineer is a bottleneck | Vacation / resignation stalls projects | Poor business continuity |
| Client preferences get lost | Client A's preferred profile, Client B's preferred glass — all in senior employees' heads | Senior staff leave = client attrition |
| Wrong orders reach the factory | Wrong drawing sent to production; materials cut before the error is found | Single-incident loss from tens to hundreds of thousands |
| Inconsistent quotes | Different sales reps quote different prices for the same project type | Client complaints + margin leakage |
| Historical quotes can't be reused | 2,000 scattered files; retrieval relies on memory | New employees can't handle orders at all |
End-to-End Workflow
Key Delivery Milestones
Day 1: Sales Rep Submits Once
| Uploaded File | System Recognition |
|---|---|
| Drawing PDF | Parsed per view per page; vector / scanned auto-distinguished |
| Window/door schedule xlsx | Row-level field extraction for each unit; tagged with source + confidence |
| Specification docx | Client default value recognition (e.g., "LowE glass for entire building") |
| Client attachment PDF | Content classification (contract / node detail / hardware brand) |
Within 5 minutes the system produces a first review card:
Project P-2026-0123: 21 window/door units. 18 units have complete fields; 3 units pending clarification. 🔴 Red 3 items (W-103 height mismatch between drawing and schedule / D-201 hardware not on NCC certified list / L2 quantity missing) 🟡 Yellow 5 items (4 units glass configuration not labeled / 1 unit opening direction unclear)
Day 1: Sales Rep Contacts Client
System automatically translates into plain-language questions, pushed to Lark:
📩 Project P-2026-0123 has 7 questions requiring client confirmation:
- W-103 height: Schedule says 1500; elevation drawing measures 1480 — which takes precedence?
- L2 master bedroom 4 windows' glass configuration not labeled — please confirm whether LowE+Argon.
- D-201 fire door hardware brand is not on the NCC certified list — please provide certification documents or change brand. ...
Sales rep forwards to client without technical jargon, with original drawing screenshot attached.
Day 2–3: Client Replies + Engineer Sign-Off
- Client returns 7 clarification answers → sales rep fills them in
- System automatically re-reconciles → 3 reds become yellows (waived with reasons) + 4 resolved
- Remaining 0 red issues → pushed to engineer for review
- Engineer completes review within 1 hour → signs off
Day 3: Quotation Decision Card
The approved structured list is automatically fed to the quotation system:
V1.xlsx rendered in parallel (identical to the existing template), ready for the sales rep to download and send to the client.
Day 4–7: Negotiation → Proforma Invoice
- Sales rep negotiates with client based on the decision card
- System validates in real time with every trial price the sales rep tries:
- Proposed price ≥ breakeven floor? ✓
- Above negotiation floor? ✓
- Deviation from historical neighbors? ±5%
- After deal closes, automatically accumulates in historical memory store: client, tier, final price, add-on combination, number of negotiation rounds
What the Agent Did / Didn't Do in This Case
What the Agent did:
- Drawing parsing + multi-view reconciliation, review health card in 5 minutes
- Client preference library auto-pulled — Client A's 30 series / Low-E glass default preferences not missed
- All mechanical work for quotation's "find historical neighbors + add-ons + floor validation" handled
- V1.xlsx identical Excel output — exactly the same as the original manual version; zero perceptible difference for clients
- Real-time "breakeven floor" validation during negotiation
- Auto-accumulation into historical memory store after deal closes
What the Agent didn't do:
- Final sign-off for red hard-blocks → engineer does this
- Negotiation strategy → sales rep does this
- Setting the cost floor → general manager sets this
- How to update client preferences → engineers manually accumulate this
- Factory BOM / internal costs → factory black box, not replicated
"Critical decisions are explainable, signable, and traceable; red hard-stops intercept wrong orders before they reach production" — this principle wasn't retrofitted; it was there from day one of design.
Actual Changes This System Brought
| Dimension | Before | After | Change |
|---|---|---|---|
| Drawing review response | Engineers review when free | Sales submits once → AI produces health card immediately | Days → 5 minutes |
| Client preferences | In senior staff's heads | Client-isolated memory store; new staff can use it too | No attrition when senior staff leave |
| Quotation output | Sales manually compares historical projects | Decision card + V1.xlsx sent directly | Hours → minutes |
| Wrong orders to factory | Occasional | Intercepted at source under hard contract | Single-incident loss avoided |
| Negotiation ceiling | Based on sales experience | Historical neighbors + floor validated in real time | Margin controllable |
| New staff independence | 1–2 years | 1–2 months | Team scales faster |
Implementation Schedule
| Week | Phase | Content |
|---|---|---|
| W1-2 | P0 (shared) | Sample governance + client preference mapping |
| W3-5 | P1 PoC (dual system parallel) | Each system re-runs 50 historical orders for validation |
| W6-8 | P2 Customization (dual system parallel) | Client rule entry + V1.xlsx template alignment |
| W9 | Integration testing | Drawing review → quotation data pipeline |
| W10-11 | Experience accumulation + company-wide go-live | Shared memory store connected + full team training |
Total duration: 10–11 weeks.
Pricing Packages
| Package | Includes | Best For |
|---|---|---|
| Drawing Review only | M1–M10 all capabilities | Customers who already have a quotation tool |
| Quotation System only | M1–M10 all capabilities | Customers not yet ready for drawing review |
| Dual System | Both full systems + hard contract + shared memory | Recommended |
| Managed Operation | Full suite + we continue maintaining | Customers who don't want to build an AI team |
For pricing details, contact info@nox-lumen.com.
FAQ
Q: Our project 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 orders as a baseline during P1.
Q: Will client preference libraries leak across clients? A: No. Each client's preferences are physically isolated at the tenant level — even logs don't cross.
Q: Can we run without LLM and use only the rules engine? A: Yes. Multi-view reconciliation, rules validation, and breakeven floor are entirely rules-engine driven. LLM is mainly used for "unstructured text understanding" (specification docx, client emails, etc.). If client inputs are 100% structured, the LLM portion can be reduced to zero.
Q: Who sets negotiation floors / breakeven floors? A: Business owners / general managers set them; the system enforces them hard. Sales rep quotes below floor outside the system → Lark alert.
Case details, contract design, and memory store architecture at docs/solutions/manufacturing/double-system.
Written by
Nox-Lumen Tech-team
Published
May 14, 2026