ZAAI System Architecture: How Horizon and Pliney Work Together

Introduction

ZAAI deploys two products: Horizon, a forecasting and decision intelligence engine, and Pliney, an agentic AI system for customer-facing operations. While they solve different problems, they are designed to work together. Horizon produces the intelligence (forecasts, scenarios, recommendations). Pliney makes that intelligence accessible through natural language and takes action on it.

This post describes the system architecture, the integration patterns, and the deployment topology in production.

Architecture Overview

The ZAAI platform consists of four layers:

Data Connector Layer — Maintains authenticated connections to enterprise systems. Each connector handles API authentication, rate limiting, schema mapping, and data normalization. Supported systems include ERP (SAP, Oracle, NetSuite, Dynamics 365, Sage), e-commerce (Shopify, WooCommerce, Magento, BigCommerce), CRM (Salesforce, HubSpot), support platforms (Zendesk, Intercom), banking APIs, and custom REST endpoints. Connectors sync data at configurable intervals (real-time, hourly, daily) depending on system capabilities and use case requirements.

Horizon Engine Layer — The forecasting and decision intelligence core. Consists of: (1) a feature engineering pipeline that transforms raw transactional data into model-ready features (seasonality decomposition, lagged variables, external signal encoding), (2) a model training and selection framework that evaluates multiple algorithms per time series and selects the best performer, (3) a hierarchical reconciliation module that ensures forecast coherence across aggregation levels using weighted least-squares, (4) a scenario simulation engine that models what-if scenarios against the trained models, and (5) an action recommendation engine that translates forecasts into operational suggestions (replenishment orders, production schedules, rebalancing actions). All components run on Google Cloud Platform using BigQuery for data warehousing and Vertex AI for model training and serving.

Pliney Agent Layer — The conversational intelligence interface. Consists of: (1) a query understanding agent that parses natural language into structured data requests, (2) an execution engine that routes queries to the appropriate data sources (including the Horizon engine), (3) a proactive alert agent that monitors operational data and surfaces anomalies, and (4) an action execution agent for customer-facing operations (order management, account changes, product recommendations). The agent layer enforces role-based access control at the query level.

Presentation Layer — Dashboards, APIs, and chat interfaces that expose Horizon outputs and Pliney interactions to end users. Includes monitoring dashboards for forecast accuracy, alert management interfaces, and embeddable chat widgets for customer-facing deployments.

Horizon-Pliney Integration

The integration between Horizon and Pliney is the core value proposition. When a user asks Pliney a predictive question ("Which products will stock out next week?"), the interaction follows this path:

1. Pliney's query understanding agent classifies the question as predictive and identifies the required data domains (inventory, forecasts, lead times)
2. The execution engine queries the ERP connector for current inventory levels and supplier lead times
3. The execution engine queries the Horizon engine for demand forecasts with confidence intervals
4. Results are joined, filtered by the user's access permissions, and formatted into a ranked risk assessment
5. The response includes data provenance (which systems were queried, data freshness timestamps) and confidence scores from the Horizon model

This integration means Pliney's answers combine three temporal perspectives: historical (what happened, from ERP data), current (what is happening, from live system connections), and predictive (what will happen, from Horizon forecasts). No other system the user interacts with provides this combined view.

Deployment Topology

A typical production deployment consists of:

• Data connectors running as managed services within the customer's cloud environment or ZAAI's GCP infrastructure, depending on data residency requirements
• Horizon engine running on GCP (BigQuery + Vertex AI) with scheduled batch processing for model training and near-real-time serving for forecast queries
• Pliney agents running as containerized services with horizontal scaling based on query volume
• Chat interface embedded in the customer's existing web or mobile application, or deployed as a standalone internal tool

Data flow is unidirectional from enterprise systems into the ZAAI platform. The platform does not write back to enterprise systems unless the customer explicitly enables action execution (e.g., automated purchase order creation).

Security Architecture

Security is implemented at every layer:

• Network: All data in transit encrypted with TLS 1.3. No data traverses public internet in customer-VPC deployments.
• Storage: All data at rest encrypted with AES-256. Customer-managed encryption keys supported.
• Access: Per-customer data isolation at the infrastructure level (separate databases, separate compute). Role-based access control enforced at both the connector layer and the agent layer.
• Privacy: No PII stored in logs or model training data. Conversational data processed in real-time without persistence. SOC 2 Type II compliant.
• Audit: Every data access, query execution, and action trigger logged with full context for compliance and debugging.

Monitoring and Operations

The platform includes operational monitoring across three dimensions:

• Model health: Forecast accuracy tracked per time series (MAPE, RMSE, bias). Automated retraining triggered when accuracy degrades below configurable thresholds.
• System health: Connector uptime, query latency, error rates, and throughput monitored with alerting. SLA tracking for response times.
• Business health: Proactive alerts for stockout risks, demand anomalies, and forecast deviations surfaced through Pliney or direct notification channels.

Conclusion

The ZAAI platform is not two separate products bolted together. It is an integrated intelligence layer where Horizon provides the computational intelligence (forecasting, simulation, optimization) and Pliney provides the interaction intelligence (natural language access, proactive alerting, action execution). The architecture is designed for enterprise deployment: data isolation, role-based access, audit logging, and infrastructure-level security. Every component is independently scalable, and the system operates in production for manufacturers, retailers, and distributors processing thousands of concurrent time series.