Federated Data & AI Marketplace & Multi-Modal Exploration

The Federated Data & AI Marketplace serves as AIX DXP’s centralized intelligence repository. Transcending the legacy limitations of disconnected data assets, the marketplace unifies business assets across Commercial, Marketing, SCM, IPD, Manufacturing, and Operations towers, gathering 229 governed data products, 100 machine learning models, 139 semantic metrics, and 130 production lakehouse tables.

The platform employs a responsive dual-pane layout: the left pane displays a multi-dimensional filtered product catalog, while the right pane features a sticky Marketplace Product Inspector supporting ad-hoc profiling, technical metadata inspection, embedded business agent simulation, and direct lakehouse access authorization.

Macro Asset Directory & Key Statistics

All assets in the marketplace adhere to standard Data Product Directory specifications and must pass AEP R1-R6 quality gates prior to publishing:

Sticky Marketplace Product Inspector Anatomy

Clicking any card in the left list (e.g., com_price_elasticity Pricing Domain) immediately populates the right-hand sticky inspector across 7 functional sections:

1. Asset Ownership & Governance Badges

• Business Tower: Explicit domain ownership (e.g., Commercial Tower, SCM Tower).
• Security & Privilege Classification: Displays regulatory tags (e.g., C4 SOX Core (Privileged Lease)), actively monitored by AEP.
• Service Level Agreement (SLA): Real-time freshness status and connection mode (e.g., Production Live Connect · Freshness SLA ≤ 15m).

2. Algorithmic Holdout Benchmark Matrix

For predictive AI data products, the inspector exposes holdout benchmark metrics verified against real production tracking samples:

• Holdout R²: Goodness-of-fit reaches 0.7200
• MAE: Mean Absolute Error bounded at 0.0167
• ROC-AUC & F1 Score: Classification certainty reaches 1.0000
• Ground Truth Sample Trace: Sample provenance and volume (e.g., n=2,985 real samples from Amazon/Walmart Tracker).

3. Enterprise 5-Modality Product Suite

AIX implements “one asset, multi-dimensional perspectives”, equipping each governed data product with 4 to 5 modal entry points:

1. Metric Explorer (Cube.js): Direct access to semantic metrics like ComKpiSnapshot and ComChannelUnitEconomics.
2. Lakehouse Table: Direct exploration of Iceberg physical tables with 4-Stage data lineage DAG.
3. ML Model: Direct access to production MLflow model cards, Feast online feature stores, and Jupyter white-box notebooks.
4. Knowledge Base: Curated domain SOPs and unstructured expert rules.
5. Event Port: 4 real-time reactive archetypes (Metric Breach, Prediction Drift, CDC Anomaly, and Infra Warning), coupled to Feishu interactive cards and Agent dispatchers.

4. Dedicated Business Agent Causal Sandboxing

Each data product embeds its domain-specific decision agent (e.g., Commercial Strategic Agent, causal attribution confidence 96.4%):

• Core Anomaly Attribution: e.g., “12 key SKUs weighted elasticity ε=-1.17; projected monthly net margin change +$774,559 (+5.29pp) for +8.2% price adjustment”.
• Risk & SLA Boundaries: Continuous monitoring against production guardrails.
• Closed-Loop Action Recommendations: e.g., “Maintain Amazon 78.4% Buy Box win rate baseline while implementing targeted inelastic SKU price increases to hedge TDI raw material +8.2% inflation”.
• Preset Inquiry Chips: 1-click prompts such as “Analyze latest lakehouse trends and drift risks?” or “What critical cross-tower dependencies exist for this asset?”.

5. Production Materialized Data Samples

The lower section displays 3 live sanitized sample rows (including skuId, channel, elasticity, recommendedDeltaPct), allowing 1-click slideout into the Entity 360 Profile.

6. Technical Metadata & Contract Specifications

Expanding the technical metadata section reveals full registry specifications:

• Lakehouse Physical Table: governed_lakehouse_mart.com_price_elasticity
• Upstream Lineage Origins: TROBS E-commerce OMS MySQL Binlog real-time stream, std_competitor_price_daily.
• Registered Contract Specs:
  • data_platform/model_products/com_price_elasticity/model_contract.yaml (ODCS Model Contract)
  • data_platform/model_products/com_price_elasticity/feature_contract.yaml (Feature Contract)
  • data_platform/model_products/com_price_elasticity/prediction_output.schema.json (Output Schema)
  • data_platform/model_products/com_price_elasticity/model_card.md (Production Model Card)
  • data_platform/notebook_products/generated/com_price_elasticity/notebook.py (Reproducible Experiment Notebook)

7. Direct Lakehouse Access Authorization (PAM & Feishu Approval)

To enable compliant data consumption, the inspector provides an integrated Request Direct Access (API / SQL) modal:

• Zero Cross-Platform Switching: Initiated in-place within DXP and linked to enterprise PAM policies.
• Time-Bound Privileged Leases: 1 Hour (Troubleshooting), 4 Hours (Standard Collaboration, Recommended), 8 Hours (Full-Day On-Duty, Upper Cap); permanent leases are strictly prohibited for SOX Core assets.
• Feishu Dual-Node Separation-of-Duties Workflow:
  • Requester: Active user account (e.g., Dennis Deng)
  • Node 1 · Line Manager: Dynamically resolved via Feishu organizational hierarchy.
  • Node 2 · Security Gatekeeper: Final authorization by Platform Security Lead (Xiong Jun · Head of AI Platform Architecture & Security).
  • Automated Issuance: Upon approval, AEP generates temporary BigQuery / REST API read-only credentials and sends in-app notifications.

8 White-Box Multi-Modal Exploration Drawers

Clicking modal tags on asset cards launches specialized full-screen inspection drawers:

1. Metric Explorer (Cube.js)

• Semantic Model Inspection: Canonical definitions, underlying SQL expressions, and dimensional slices (Channel, Region, SKU).
• SQL Compilation: Instant compilation to Google BigQuery SQL for verification in Cloud Console.

2. ML Model Card (MLflow / TreeSHAP)

• Lifecycle Records: Model versions, training distributions, and evaluation metrics (AUC-ROC 0.942, PR-AUC 0.915, RMSE 0.042).
• TreeSHAP Attribution: Global and local feature importance rankings illustrating pricing, competitor, and inventory sensitivities.

3. Feast Feature Store Explorer

• Unified Online/Offline Features: Low-latency online store (Redis, latency (< 5ms)) and offline lakehouse store (BigQuery Parquet).
• Entities & Feature Views: Instant retrieval of rolling window features across sku_id, customer_id, and warehouse_id.

4. Interactive Jupyter Notebooks

• Code as an Asset: Pre-packaged, reproducible Python notebooks containing feature engineering and LightGBM hyperparameter search code.
• Export to Colab / Vertex AI: 1-click cloud sync for extended experiments.

5. Data Lineage DAG Explorer

• End-to-End Tracing: Interactive DAG showing source (Amazon SP-API / Shopify Webhook) $\rightarrow$ Cloud Run CDC $\rightarrow$ Bronze $\rightarrow$ Silver $\rightarrow$ Gold $\rightarrow$ BI Engine $\rightarrow$ Cockpit.
• Column-Level Lineage: Granular column-level tracking for upstream schema change impact analysis.

Multi-Dimensional Catalog Filtering

• By Control Tower: All (229), Commercial (68), Marketing (34), SCM (38), IPD (33), Manufacturing (25), Operations & Finance (31).
• By Modality: All (229), AI Models (100), Domain Data Products (129).
• By Governance Tier: Tier-1 Mission Critical (SLA 99.99%) vs. Tier-2 Operational.
• By Quality Score: Ordered by ODCS Quality Score (90 - 100).
• Full-Text Semantic Search: Fuzzy search across titles, descriptions, physical tables, and column comments.