MarketPulse
All chapters

03 — Phase 1A: The Localization Matrix

Goal: decide what stays global (build once, serve everywhere) vs. localized (adapt per market) so we maximize insight quality while minimizing engineering bloat.

The governing principle: localize the edge, standardize the engine

A trend-intelligence platform is a pipeline: collect data → process with ML → surface insights → generate & screen concepts → deliver in a workflow. The intelligence (ML models, ranking, generative concepting) is universal. What's local is the data going in, the language it's in, the categories it maps to, the laws around it, and the workflow it plugs into. So we standardize the middle and localize the two ends.

        GLOBAL (build once)                          LOCALIZED (adapt per market)
  ┌──────────────────────────────┐        ┌────────────────────────────────────────┐
  IN →  Data ingestion framework   →  ML/NLP core  →  Insight & concept engine  → OUT
        ▲ localized data SOURCES                                    localized WORKFLOW ▼
        localized LANGUAGE/NLP packs                                localized COMPLIANCE
        localized TAXONOMY/categories                               localized GTM/support

The matrix

Each capability is scored on two axes — does localizing it materially increase customer value? and how expensive/ongoing is it to localize? That yields a verdict.

CapabilityLocalize?WhyLocalization costVerdict
ML/forecasting core, ranking algorithmsGlobalMath is market-agnosticn/aStandardize 100%
Concept Genie generative engine (model)GlobalSame model; only prompts/data differLowStandardize; localize inputs
Data ingestion frameworkGlobalOne pipeline, many connectorsn/aStandardize
Data sources (local e-com, menus, retailers, social)LocalizeInsight quality = local data depth; this is the #1 value driverHigh, ongoingLocalize (config, not code)
Language / NLP packs (already 18 langs)LocalizeTrends are expressed in local languageMedium, ongoingLocalize via language packs
Category taxonomy (e.g., "snacking", "ayurvedic", "halal", "wellness")LocalizeCategories are culturally specificMediumLocalize as config/data
Brand/competitor reference setsLocalizeLocal brands differMediumLocalize as data
Compliance / data residency / privacy (GDPR, PDPA, DPDP)LocalizeLegal requirement to sell enterpriseMedium, ongoingLocalize by region (infra + policy)
Enterprise workflow integrations (their PLM/innovation stack, SSO)Localize (tiered)Big-enterprise buyers demand it; SMBs don'tHigh per integrationLocalize only at Enterprise tier
Reporting / export templates, currency, unitsLocalizeLow effort, high polishLowLocalize (i18n)
UI translationLocalize (light)Expected but not the value driverLowLocalize last; English-first acceptable in many B2B markets
Pricing, contracting, invoicing currencyLocalizeWTP & finance norms differLow-MedLocalize commercially (doc 05)
Onboarding & support content/hoursLocalize (tiered)Service overhead driverMediumLocalize by tier + region (doc 06)

Config-not-code: the bloat firewall

The single most important engineering rule for avoiding bloat: localization must be data/configuration, never a code fork. Build a market pack abstraction:

A Market Pack = { language model(s), data-source connectors, category taxonomy, brand reference set, compliance profile, currency/format, support config }. Adding a market = shipping a new pack, not branching the product.

This keeps ~80% of the codebase shared (Assumption #2), turns "enter a market" from an engineering project into a configurable, repeatable operation, and is exactly what the automated workflow provisions.

The trade-off, made visual (cost-to-serve vs. revenue-uplift)

  high │  Compliance/        │  Local data sources
 rev   │  residency          │  Language/NLP packs
 uplift│  (do it — required) │  Taxonomy  (DO FIRST — high ROI)
 ──────┼─────────────────────┼──────────────────────────────
  low  │  UI full translation│  Deep per-customer workflow
 rev   │  (defer)            │  integrations (Enterprise-only,
 uplift│                     │   charge for it)
       └─────────────────────┴──────────────────────────────
            low loc. cost            high loc. cost

Do-first quadrant (high uplift, manageable cost): local data sources, language, taxonomy. Charge-for-it quadrant (high uplift, high cost): deep integrations → gate behind Enterprise tier and price as services. Defer quadrant: full UI translation. Required regardless: compliance/residency — it's table stakes to even sell.

Machine-readable version: frameworks/localization-matrix.csv.