Anatomy of a Corporate AI Platform: How an LLM gateway, its admin panel, and its user portal are built inside

This is not a book about prompting or adoption. It is the technical documentation of the internal architecture of a real corporate AI gateway, read from the code: the stage pipeline, the multi-provider router, caching, compression, quotas, cost, content security, audit, governance, MCP federation, the admin panel, and the user portal - piece by piece, with diagrams and a reference to the exact repository file.

The case study is N7xGateway: FastAPI + SQLAlchemy 2, MySQL, Redis, Celery, 7+ LLM providers (Anthropic, OpenAI, Azure, Gemini, Bedrock, Ollama, LM Studio), and a React panel. Its references are anonymized: real mechanisms without exposing proprietary structures or client data.

34 technical chapters across 16 parts, plus 6 appendices, covering the full anatomy of the platform:

• The engine - The PipelineRunner and its 14 stages over a mutable context: auth, input security, PII, cache, quota, routing, enrich, trim, output security, audit. Bidirectional SSE streaming with filtering and audit inside the generator.
• Routing - RouterService by purpose, smart model selection without ML, data residency. DeploymentRouter with health, cooldown, and fallback chains. Adapters for 7+ providers with payload normalization and tool calling.
• Efficiency - Exact, dual, and semantic caching; native prompt cache. Context compression with BM25 scoring and stubbing. Precise token counting.
• Beyond chat - Embeddings, images, and audio with cost, PII, and audit; backend tools (web search and fetch with SSRF defense, document extraction); queued inference with Celery and RAG.
• Cost and control - Tenant isolation, windowed quotas, per-plan entitlements. Decimal pricing, reconciliation, and ROI. Rate limiting with sliding window, token bucket, and circuit breakers.
• Security and trust - JWT with device binding, SSO/SCIM/MFA, AES-256-GCM and break-glass. Jailbreak, MSJ, DLP, PII. Classification, guardrails, firewall. Append-only audit chained with SHA-256 and GDPR pseudonymization. Governance and compliance (GDPR, AI Act, DSA).
• MCP, panel, and portal - Per-tenant MCP federation and the n7x-mcp server. Three React SPAs with realtime over WebSocket: the admin panel and the self-service user portal.
• Data and infrastructure - The data model (127 tables, append-only, encryption). Docker Compose, hardened Nginx, Celery-based deployment. Plus glossary and catalog appendices.

• What problem it solves - the design decision and its trade-off.
• How it works - the flow, the data structures, and the invariants, with prose and diagrams.
• How it is made secure - what breaks if you build it naively and what control prevents it, showing the control and never the exploit.

Platform architects, backend engineers, security and compliance leads, frontend developers, and CTOs who want to understand the mechanics before approving the architecture. After reading it, you can deliver a technical training session on the architecture of an AI gateway the next day. That is the bar.

Anatomy of a Corporate AI Platform keeps the form opened by Anatomy of an AI Agent: the object of study is the machine itself - here, the platform that governs the corporate use of models.

Carlos Pérez González - AI solutions architect. OSCE, OSCP, OSWE, OSEP, CREST.
Juan C. Montes - Cybersecurity architect. GCFA, GREM. Published in PHRACK #65.