Skip to content

AI & Models · production tested · 19 components

AI Application

From idea to production. An AI application inside a real organisation, drawn end to end, so you can read it as data, networking, security, identity, integration, compute, operations and cost.

Pick a lens to read it as one discipline. Select a component to see what it means and what is written about it.

StackDetail

How to read this

Each box is a role. Pick a lens to see the system as a networking, data, security or identity problem. Pick a stack, or mix stacks per component, to see the product that fills each role.

Request
Async or batch
Telemetry and audit
Trust boundary
Cross-platform seam
Seam not drawn as a line

Adds identity carried end to end, secrets in a vault, an audit record for every call, and real systems of record behind the tools.

Components

What an AI application is

Strip away the demos and an AI application is a normal enterprise system with three new parts: a model you do not control, retrieval that feeds it context, and tools it is allowed to call. Everything around those three is engineering you already know: networks, databases, identity, integration, operations and cost.

That is why the map above has lenses. Pick one and the same system turns into a networking problem, a data problem or an identity problem, with the components that matter for that discipline lit up.

The shape that survives production

Edge and API gateway own the way in: TLS, WAF, authentication, rate limits.
Agent decides what happens next. In most systems it is a supervisor with a small set of specialist workers.
LLM gateway is the only way out to a model: routing, budgets, redaction and one place to trace every call.
Retrieval supplies the knowledge the model was never trained on, filtered by who is asking.
Ingestion keeps that knowledge fresh from the systems that own it.
Tools are typed MCP servers with explicit permissions in front of real enterprise APIs.

Where teams get hurt

Observability arrives too late, evaluation is a spreadsheet, identity stops at the web app, and the deployment story is "it runs on my laptop". The guides attached to each component address these in order.

Node by node

Web App

Frontend3 linked

Understand

ArticleWhy a capable model is still not a product

Implement

Build logBuild log: shipping Unseen UI to npm

Use

ToolUnseen UI

Open Web App on the map

Identity

SSO · OBO2 linked

Understand

ExplainerYour agent should not be a superuser

Operate

ChecklistSecurity review checklist for an AI feature

Open Identity on the map

API Gateway

AuthN · limits4 linked

Understand

ExplainerYou are already building a control plane, badly

Operate

ArticleWhat an AI gateway actually costs to run

Use

ToolOpenTelemetry

Reference

ReferenceOpenTelemetry Documentation

Open API Gateway on the map

Secrets

Vault · KMS1 linked

Operate

ChecklistSecurity review checklist for an AI feature

Open Secrets on the map

Agent

Orchestration34 linked

Understand

ExplainerWhy complex AI questions need decomposition
ExplainerWhy reflection only matters when tied to evidence
ExplainerWhy agents need outcomes, boundaries, and stop conditions
ExplainerWhy agent behavior is a loop, not a single prompt
ExplainerWhy planning reduces wasted AI actions
19 more on the map

Implement

Decide

ComparisonWhen to use a chatbot, workflow, or agent
ComparisonMost teams reaching for multiple agents do not need them
ComparisonMost agents are workflows wearing a costume
ComparisonLangGraph vs the OpenAI Agents SDK

Use

ToolOpenAI Agents SDK
ToolLangGraph

Reference

ReferenceBuilding Effective Agents

Open Agent on the map

Audit

Every call7 linked

Understand

ExplainerWhy multi-agent systems need decision rules
ExplainerWhy governance belongs in the architecture
ExplainerYou cannot instruct a model into data protection
Architecture patternPattern: the outbox for agent actions

Operate

ChecklistSecurity review checklist for an AI feature
ChecklistWhen to bring in a compliance review
ArticleWhat an AI gateway actually costs to run

Open Audit on the map

LLM Gateway

Route · budget10 linked

Understand

ExplainerWhy tokenization quietly affects cost, limits, and reliability
ExplainerWhy guardrails should enable safe action
ExplainerWhy autonomy needs budgets
ExplainerSafety is an architecture decision, not a moderation setting
ExplainerUntrusted text does not get to give orders
5 more on the map

Open LLM Gateway on the map

Observability

Traces · evals18 linked

Understand

ExplainerWhy RAG must be evaluated in parts
ExplainerWhy observability is not optional for AI systems
ExplainerWhy evaluating the model is not enough
ExplainerThe 3am agent run that nobody is watching
ExplainerYou cannot run RAG on user complaints
11 more on the map

Use

ToolOpenTelemetry

Reference

ReferenceOpenTelemetry Documentation

Open Observability on the map

Model

Provider API13 linked

Understand

ExplainerWhy model choice is rarely the first production AI problem
ExplainerWhy LLMs feel intelligent but still generate one step at a time
ExplainerWhy tokenization quietly affects cost, limits, and reliability
ExplainerWhy more context can make an AI system worse
ExplainerWhy vague prompts become vague systems
3 more on the map

Implement

ChecklistWorking with Claude, practices that hold up

Operate

ChecklistWhen to bring in a compliance review

Use

ToolLiteLLM
ToolGuardrails

Reference

ReferenceAttention Is All You Need

Open Model on the map

Sources

Docs · tickets · DBs5 linked

Understand

ExplainerWhy retrieval quality starts before search
ExplainerWhy production knowledge often lives in systems, not PDFs
ExplainerWhy freshness is part of correctness
ExplainerYour vector index does not know who is allowed to read it

Operate

ChecklistWhen to bring in a compliance review

Open Sources on the map

Ingestion

Chunk · embed8 linked

Understand

ExplainerWhy retrieval quality starts before search
ExplainerWhy chunk boundaries shape answer quality
ExplainerWhy embeddings are useful and easy to overtrust
ExplainerWhy evidence is becoming multimodal
ExplainerWhy freshness is part of correctness
3 more on the map

Open Ingestion on the map

Vector Index

Embeddings · hybrid4 linked

Understand

ExplainerWhy embeddings are useful and easy to overtrust
ExplainerWhy vector stores are infrastructure, not magic memory
ExplainerWhy RAG needs both broad context and precise evidence
ExplainerWhy some answers live in relationships, not documents

Open Vector Index on the map

Retrieval

Hybrid · rerank23 linked

Understand

ExplainerWhy confident AI answers still need evidence
ExplainerWhy RAG is an evidence design problem, not a buzzword
ExplainerWhy keyword search and semantic search both matter
ExplainerWhy hybrid search is often the practical default
ExplainerWhy relevant data can still be unauthorized data
12 more on the map

Operate

ChecklistRAG production-readiness checklist
Failure storyThe retrieval cache that served stale policies

Use

Reference

ReferenceRAG at Scale: Architecture and Lessons

Open Retrieval on the map

Tools

MCP servers20 linked

Understand

ExplainerWhy production AI needs structured outputs
ExplainerWhy production knowledge often lives in systems, not PDFs
ExplainerWhy tool access is where AI becomes operational risk
ExplainerWhy handoffs need contracts
ExplainerWhy retrieved content must stay untrusted
9 more on the map

Decide

Operate

ChecklistSecurity review checklist for an AI feature

Use

ToolInstructor

Reference

ReferenceModel Context Protocol
ReferenceOWASP Top 10 for LLM Applications

Open Tools on the map

Enterprise APIs

ERP · CRM · ITSM2 linked

Understand

ExplainerAn approve button is not human oversight

Decide

ComparisonWhen MCP earns its overhead, and when a direct API is better

Open Enterprise APIs on the map

Memory

Session · long-term6 linked

Understand

ExplainerWhy agent memory is not one database
ExplainerWhy shared state needs structure and attribution
ExplainerWhy shared memory becomes a permission problem
ExplainerWhy handoffs need contracts
ExplainerFour kinds of agent state, and only one of them is memory
1 more on the map

Open Memory on the map

Human review

Approval queue3 linked

Understand

ExplainerWhere human review actually belongs in AI workflows
ExplainerWhy multi-agent systems need decision rules
ExplainerAn approve button is not human oversight

Open Human review on the map

On this map