Equinox Intelligent Platform — Architecture Overview

A layered view of the Equinox Intelligent Platform — from Unified Chat and the agentic runtime down through tools, AWS AI services, and storage — showing how an intelligent commerce request flows end to end.

This diagram presents the end-to-end runtime and tooling architecture of the Infosys Equinox Intelligent Commerce Suite (ICS) — a multi-tenant, agent-driven platform built on AWS. It is organized as a top-to-bottom stack of layers, showing how a user request flows from the interface down through the agentic runtime, tools, AI services, and storage.

Entry & Access. At the top, Users/Admins and External Systems (webhooks, events, APIs) reach the platform through AWS Cognito, which handles authentication and multi-tenant isolation with a pool-per-organization model scoped across Store, Business, and Global levels.

Frontend. A React 18 + Vite single-page app provides two surfaces: a Dashboard (landing view for resources, executions, and health) and a Unified Chat / Conversation Manager that interprets user intent and routes it to the right agent or workflow, streaming responses via AG-UI SSE.

Backend Services. A FastAPI (Python) layer organized as api → services → repositories, containing the core control-plane modules: tenancy (identity, RBAC, org registry, provisioning), agent-studio (the LangGraph agent runtime, workflow engine, and NL authoring), connectors (a domain-verb capability router over MCP/REST/GraphQL/gRPC), and track tools for Prediction (ML models), Knowledge (RAG / institutional brain), and Language (NLP/fine-tuned SLMs).

Build, Manage & Operate Layer. A user-facing control layer, reachable from Unified Chat, for authoring agents and workflows, managing platform resources (Agents, MCP/Tools, Guardrails, Skills, Prompts, Workflows), and running domain operations.

Agentic Engine Layer (Runtime). The orchestration core: Triggers (manual, chat, webhook, event, schedule) feed a Supervisor/Orchestrator that routes intent to specialist agents, a Workflow & Graph Engine (LangGraph, multi-agent, human-in-the-loop), a Guardrail Pipeline (pre/post-call checks, PII masking), and Execution History. Temporal (flagged NEW) adds durable workflow orchestration with retries and interrupt/resume.

Tool Layer. Defines how agents act — agent-attached MCP tools (bound to a specific agent, including private/third-party MCP servers) and platform tools (native controls available to every agent, such as memory and resource lookup).

AI & Secrets Layer (AWS managed). AWS Bedrock (live) for foundation-model access (Claude Sonnet 4.6, Titan, Nova, Mistral, Qwen), a LiteLLM Gateway (NEW) providing a unified multi-provider LLM proxy with cost governance and fallback, AWS SageMaker for ML training/inference and SLM fine-tuning, and AWS Secrets Manager for per-org key management and runtime secret injection.

Storage Layer (AWS + Atlas). Polyglot persistence: DynamoDB (control and data plane), OpenSearch Serverless (vector search for KB/memory and log analytics), Neptune Analytics (GraphRAG / knowledge graph), MongoDB Atlas (agents, workflows, prompts, executions), and Redis (caching, checkpoints, session state).

Shared Library. A common library compiled into every service (context/tenant resolution, base repositories, structured logging, audit, guardrail middleware, LLM/secrets providers) — infrastructure only, no business logic.

Request Flow. A numbered path across the bottom summarizes the lifecycle: (1) Dashboard entry → (2) Unified Chat interprets intent → (3) build/select agents & workflows → (4) engine orchestrates & runs → (5) agent invokes tools → (6) response surfaced back in Unified Chat.

Revision History
2026-08-12 | JP – Created the page and added the content.