AI Architecture

Production AI systems are not single components - they are stacks. A retrieval layer grounds the model in your data, an agent layer decides what to do with that grounding, an orchestration layer coordinates multiple agents and tools, and a protocol layer like MCP wires those tools together in a standard way. Fine-tuning and guardrails sit across all of it: fine-tuning changes what the model knows and how it behaves at the weight level, guardrails constrain what it's allowed to do at runtime.

This section is the reference architecture for that stack. Each page below covers one layer or technique in depth - read them together to understand how a request moves from a user question to a grounded, tool-using, safety-checked answer.

How These Pieces Fit Together

Data flows left to right through retrieval, agents, and tool integration. Fine-tuning and guardrails (dashed lines) are cross-cutting - they apply at multiple layers rather than occupying one place in the pipeline.

In This Section

• RAG (Retrieval-Augmented Generation) - production RAG patterns: chunking strategies, retrieval architectures, re-ranking, evaluation, and failure modes.
• AI Agents - AI agent architectures: tool use, function calling, ReAct pattern, memory systems, and production deployment patterns.
• Multi-Agent Systems - multi-agent orchestration patterns: sequential, concurrent, handoff, group chat, and magentic topologies for production AI systems.
• Autonomous AI - levels of AI autonomy, safety guardrails, human-in-the-loop patterns, and the path from supervised to self-directed agents.
• Workflow Automation - triggers, orchestration, multi-agent task pipelines, human-in-the-loop, and the operational scaffolding that wraps an agentic workflow in production.
• MCP (Model Context Protocol) - building MCP servers, connecting tools to LLMs, transport mechanisms, and production deployment patterns.
• Text-to-SQL - natural language to database queries: schema grounding, query generation, validation, and production deployment.
• Fine-tuning - when and how to fine-tune models vs prompt engineering vs RAG: trade-offs, workflows, and production patterns.
• Guardrails & Safety - content filtering, prompt injection defence, PII protection, and responsible AI patterns for production systems.
• OCR + Document Processing - production document intelligence: OCR, layout detection, table extraction, and LLM-powered structured extraction with confidence scoring.
• The AI Layer - building the reasoning layer on top of your data warehouse: semantic definitions, verified examples, and metric trees that make AI-powered self-service analytics trustworthy.