Ontology-Driven Agentic AI System — Merchandising & Data Science
Microsoft + Anthropic · Frugal AI · Modular · Autonomous · No interactive user at runtime · V5
🧠 Anthropic Claude ☁ Microsoft Azure ⚡ Frugal AI · Token Economy 🔗 Enterprise Integration
Data Sources
Input Layer
Decision Processing
Shared Knowledge
Builder Agents
Artifacts Library
Intelligence Layer
Agent Library
ML / Analytics
MCP Framework
Frugal AI
Output
Integration
FLOW 8 steps · ontology-driven · orchestration determines complexity · agent auto-selects LLM/model · write-back to enterprise 1 Data SourcesExternal+Enterprise 2 Input LayerBiz+Data Ontology 3 KnowledgeFactory 4 BuilderAgents 5 ArtifactsLibrary+KG 6 IntelligenceLayer 7 OutputLayer 8 IntegrationWrite-back ⬛ FRUGAL AI · CROSS-CUTTING ⬛ MCP · CROSS-CUTTING SHARED KNOWLEDGE Tenant-agnostic · preloaded ● Retail KPI base ● Ontology patterns ↳ KPI definitions · benchmarks · retail norms feeds factory Decision Processing → combined outputs → Output Layer ← Intelligence Layer returns results · Decision Processing combines & delivers ↑ results DATA SOURCES 3 integration surfaces · symmetric read & write-back EXTERNAL SYSTEMS connected via API · write-back capable ↔ API · event-driven · read & write-back Relex Anaplan Blue Yonder + and others CLOUD SYSTEMS platforms · storage · processing AWS Azure GCP Snowflake Databricks DATA STORAGE enterprise systems (ERP · operational) SAP Oracle MS Dynamics Workday DATA ENTITIES · queried at runtime queried at runtime · shared across all sources sales_transactions product_catalogue pricing_history promotion_events inventory_positions store_hierarchy customer_segments competitor_prices seasonal_calendar merch_financials range_attributes supplier_data merch_kpis_actuals cluster_attributes KPI actuals → Shared Knowledge / KPI base 🧠 Intelligence overlay on all data entities KPI targets & definitions → Shared Knowledge external + cloud APIs → → C1 Input storage & entities → LAYER 1 · INPUT LAYER Ontology-driven · setup phase Business Ontology User & Business Team Defined · Setup GOALS & STRATEGY Goals & KPIs Business rules Guardrails CONSTRAINTS Financial targets Brand / trade rules Seasonal calendars doc · sheet · slide · text → parsed by Biz Ontology Builder Merch KPIs NOT here → Retail KPI base KPI targets from Shared Knowledge ↕ Created from Data Sources ← Data Ontology SCHEMA & STRUCTURE Schema & data dictionary Entity relationships Metadata & lineage Semantic types SYSTEM MAPPINGS Relex schema mapping Anaplan model structure Blue Yonder data contracts ERP table definitions → parsed by Data Ontology Builder → used to construct Knowledge Graph business context → data schemas ↑ DO→C2 LAYER 2 · DECISION PROCESSING 01 Read Knowledge & Invoke Builders KPI base + ontologies → invoke builders (setup only) builds + caches artifacts & KG ← reads Artifacts Library & KG 02 Plan & Invoke Intelligence Layer Orchestration determines task complexity Routes: agent · ML · analytics → Frugal AI routing → instantiates via MCP → parallel execution possible 03 Combine & Deliver Orchestrate all outputs Synthesise component results → Output Layer ↓ via flow corridor below ORCHESTRATION NOTE Orchestration determines complexity Agents auto-select LLM / model invoke builders → PRIMARY ← artifacts ← KG → Factory LAYER 3 · BUILDER AGENTS & ARTIFACTS LIBRARY BUILDER AGENTS · SETUP PHASE Parse ontologies → produce artifacts Biz Ontology Builder → YAML · guardrails → system prompts Data Ontology Builder → schema maps → KG construction ← Business Ontology ← Data Ontology produces ↓ ARTIFACTS LIBRARY Own library · cached runtime YAML configs Guardrails Schema maps Column defs System prompts KNOWLEDGE GRAPH Own library · entity-aware Client-specific Entity relations Causal links Semantic graph Context vectors Consumed by Factory (C2) ← reads at every runtime request SETUP PHASE (one-time) Builder agents invoked → produce artifacts Artifacts & KG cached · versioned RUNTIME PHASE (per request) Factory reads cached artifacts & KG Intelligence Layer invoked per task KG refreshed on data change Factory Step 02 → invokes Intelligence Layer at runtime ↓ invoke → Layer 4 LAYER 4 · INTELLIGENCE LAYER 3 parallel libraries · agent auto-selects model · governed by Frugal AI Artifacts Library & KG consumed by Factory (C2) · passed as context payload to agents at invocation FRUGAL AI FRAMEWORK LLM auto-selection · agents choose capability tier per task complexity, cost & speed TOKEN OPTIMISATION prompt compression · context scoping · output caching · response length · result reuse AGENT LIBRARY LLM-powered · tool-use via MCP ACTIVE AGENTS Pricing Promotion Demand Forecast Price Elasticity Trend Analysis Assortment AVAILABLE IN LIBRARY Inventory replen. Range planning Space planning ● Active ○ Available AUTO-SELECTS LLM/MODEL task complexity · cost per Frugal AI banner ↑ ML & MODEL LIBRARY Statistical · ML · Optimisation FORECASTING SARIMA / SARIMAX Prophet XGBoost / LightGBM REGRESSION/CAUSAL OLS / Ridge / Lasso Double ML / DML ETS / Holt-Winters OPTIMISATION Constraint optimiser K-Means clustering SHAP / explainability NLP · ANOMALY Embedding models Anomaly detection DIAGNOSTIC Causal inference Attribution analysis DESCRIPTIVE ANALYTICS Token-free · no LLM required Aggregations & pivots KPI dashboards Trend reports Rule-based alerts Threshold monitoring Data slicing Ad-hoc comparisons ↑ Zero LLM token cost Max cost efficiency For complex analytics: → Agent Library → ML Model Library Simple cases only No LLM instantiation needed MCP TOOLING FRAMEWORK · CROSS-CUTTING Connects: Artifacts Library · Agent Library · ML Library · Descriptive Analytics routes all tool calls · authentication · typed results returned to Factory ML model tools Data warehouse tools Rule engine tools External API tools PER-TASK EXECUTION PATTERN Orchestration selects library type → agent auto-selects LLM tier (Frugal AI) → MCP routes tool calls → typed JSON Factory combines all component results → Output Layer · results cached where applicable results → LAYER 5 · OUTPUT LAYER AI decisions · plans · insights DECISION OUTPUTS Pricing Optimised prices Scenario analysis Promotion Discounts & eligibility Uplift forecasts Demand fcst SKU-level projections Trend + KPI signals Inventory Replenishment signals Allocation decisions Financial OTB budget recs Margin targets Assortment Range decisions Whitespace gaps Space Planning Fixture allocation Category adjacency Supplier & Sourcing Reorder signals Lead time optimisation Customer & Loyalty Segment offers Churn risk signals Markdown & EoL Clearance timing Discount depth by SKU expected system capability · operationalise AI outputs WRITE-BACK DESTINATIONS close the loop · operationalise AI outputs EXPECTED SYSTEM CAPABILITY EXTERNAL SYS Relex · Anaplan Blue Yonder · o9 forecasts · plans OTB · write-back CLOUD SYSTEMS AWS · Azure GCP · Snowflake predictions · audit ML feedback DATA STORAGE SAP · Oracle MS Dynamics pricing · POs master data ↩ closes loop · writes back to source systems refreshes Layer 1 inputs · next planning cycle 1 2 3 4 5 6 7 8 9 10 11 12 ① Data Sources · ② Input Layer · ③ Knowledge Factory · ④⑤ Builder Agents + Artifacts & KG · ⑥ Intelligence Layer · Agent Library · ML & Model Library · Descriptive Analytics · Frugal AI banner · ⑦⑧ Output + Integration