Case study: AI quotation system

Background: Real overseas window engineering rollout for Nox-Lumen Mfg. Every plant differs; use this page as reference only. Follow the four-phase delivery playbook documented on the main manufacturing hub.

Elevator pitch

Deliver a project-level blended $/m² recommendation (add-ons included) that never goes below floor margin while respecting statistically similar historic deals. Turns ~2 000 stray quotes into searchable institutional memory.

The assistant never substitutes the salesperson’s judgement and never reconstructs factory BOMs—factory CNY totals are opaque inputs only.

End-to-end flow

Sales       Upload factory BOM xlsx + quoting notes docx (+ optional PDF metadata)
↓
Platform    Parse → tier tagging → similarity search over archived projects
↓
Platform    Compute break-even floor • recommended price • negotiation slack • add-on rollup
↓
Platform    Produce “quotation decision cards” (+ neighbour evidence)
↓
Approvers   Tier-based routing notifications (Feishu / WeCom)
↓
Platform    Approved → render V1.xlsx (matches today’s templates) → download for customer
↓
Platform    Persist quote, negotiation deltas, wins for future retrieval

Core capabilities

1. Ingest quoting collateral

File	Handling
Factory quote xlsx (CNY)	Table detection + row extraction (series, spec, qty, unit price)
Narrative docx	System glazing stack, variants, exclusions
Attachment PDF	Optional geo / scope hints

2. Tiering engine

Total glazed area tiers ensure XS jobs don’t train XL pricing (five buckets by default):

Tier	Range	Typical job
XS	Under 50 m²	Single-house retrofit
S	50–200 m²	Compact residential
M	200–800 m²	Standard apartments / rows
L	800–3 000 m²	Large homes / light commercial
XL	> 3 000 m²	Campus / podium-scale

Tier edges are adjustable per rollout.

3. Historic memory corpus

Attribute	Behaviour
Capacity	Bulk import ~2 000 historical quotes + incremental new deals
Retrieval	Semantic match (geo, façade type, glazing brand, hardware) + tier filtering
Neighbours	Return top 3–5 neighbours with blended $/m² plus deal outcomes
Influence	Anchor the recommendation via neighbour medians ± tier / customer uplift factors

4. Pricing stack

recommendation = max(
    break_even_floor,
    neighbour $/m² × tier_factor × customer_factor × uplift_factor
)

Every numeric output links back to factory spreadsheet rows, tier, coefficient provenance, and neighbour IDs.

5. No-loss sentinel

Item	Detail
Formula	`floor = factory_cost × (1 + minimum_margin)`
Margin table	Tier & customer archetype aware—business stewards edit centrally
Hard rule	Recommendation & negotiation slack never dip below floor
Escalations	Alerts if reps quote beneath floor offline

6. Add-on synthesiser

Screen cloth, trims, auxiliary frames, hardware upgrades, special finishes—normalized naming, priced, rolled into blended $/m² plus line-item appendix.

7. Excel renderers

Output	Audience
V1.xlsx	Customer-facing first quote
Decision card UI	Sales & leadership introspection
PI.xlsx	Post-negotiation instrument

Baseline customer templates ported within ~one week each engagement.

8. Approval matrix

Trigger	Typical chain
XS/S + priced above reco	Solo sales
M	Sales lead
L/XL	Sales lead + director
Negotiation scraping floor (below floor×1.05)	Director + Finance
At floor itself	Director + executive

Theme	Guarantee	Mechanism
No-loss maths	Outputs never violate floor margin	Server-enforced inequalities
Human-in-loop	Recommendation is advisory	Sales confirms final envelope
No fake BOM cloning	Factories retain costing IP	Imports are opaque aggregates
Explainability	Every figure traceable	Immutable audit breadcrumbs
Data residency	Assets never mingle across tenants	Dedicated MinIO scopes

Explicitly out of scope (MVP)

Wish	Reason	When to revisit
Auto CAD ingestion	overlaps drawing-review module	Evaluate bundle roadmap
Rebuilding factory costing	violates commercial boundary	Never primary goal
Auto PI legal prose	Legal owns contracts	Separate legal track

Implementation snapshot

Phase	Duration	Goal	Acceptance
P0 data hygiene	~2 wks	Harvest 30–50 exemplar deals + tier thresholds	Similarity QA metrics
P1 parsers + tiers	~2 wks	Parsing + searchable archive	Upload→neighbour drill
P2 reco + outputs	~3 wks	Engines + Excel render	Sales can produce V1 end-to-end
P3 governance	~1 wk	Routing + BOM revisions	Fixtures pass regressions
P4 adaptive learning	~2 wks	Coefficients + reporting	Monthly report ship

Rough calendar 8–10 weeks.

Full SoW → info@nox-lumen.com.