Maps
Production AI Platform Map
The twelve-responsibility map from use cases to ownership.
Field guide · v0.1
Production AI Platform Handbook
From API key to platform.
Production AI is not a model. It is a platform. This is a practical map for teams building LLM, STT, embeddings and agent systems in production: inference, routing, cache, evals, guardrails, observability, cost and ownership.
14
chapters and maps
12
platform layers
5
content formats
3
tools
4
role tracks
API key → platform
01
Product use cases
02
AI Gateway
03
Provider strategy
04
Execution-path routing
05
Inference runtime
06
Caching
How to use the handbook
Do not read it linearly. Start from the platform problem in front of you.
Need the big picture?
Open the platform map and locate the missing responsibility.
Need an executive framing?
Use the maturity model and MaaS vs self-hosted strategy chapter.
Cost or latency is drifting?
Start with inference economics and prefix cache.
Quality is unstable?
Use AI Quality Gate, observability and ownership chapters.
Start by role
Pick the path closest to your current responsibility, then use the catalog below to filter materials.
AI Platform Lead
Ownership, platform map, quality and economics.
CTO / Head
Strategy, maturity and operating model.
Backend to AI
Gateway, observability, tools and production contracts.
MLOps / Inference
Serving, STT, embeddings, cache and capacity.
Handbook navigator
Search the first release by role, format or topic, then mark what you have read. Progress stays in this browser only.
Progress
0/22
read
0
bookmarked
Local progress only. No account, tracking or backend.
Search
Formats
Role tracks
Maps
AI Platform Maturity Model
Seven levels from one API key to AI-native operations.
Chapters
Start Here
Choose the first chapter by the platform problem in front of you.
Chapters
Manifesto
Why production AI should be treated as a platform, not a model choice.
Chapters
MaaS vs Self-hosted
A strategy decision, not a religious argument.
Chapters
AI Gateway
The control point for access, quotas, routing, policy and cost.
Chapters
Semantic Router
Route execution path, not only model names: direct, RAG, agentic or human review.
Chapters
Inference Runtime
Serving decisions around LLM, STT, embeddings and rerankers.
Chapters
STT Serving
Treat speech-to-text as a first-class production workload.
Chapters
Embeddings Serving
Embeddings and rerankers as a quality, latency and lifecycle layer.
Chapters
Inference Economics
Cost per accepted outcome, not raw token math.
Chapters
Prefix Cache
How prompt shape, tool schemas and routing decide cache behavior.
Checklists
AI Quality Gate
A rollout loop that prevents silent quality degradation.
Checklists
LLM Observability Checklist
Minimum telemetry for model, prompt, cost, latency and outcome.
Chapters
Guardrails and Security
Policies, telemetry, fallback and ownership, not one magic library.
Chapters
Ownership and Operating Model
Who owns quality, cost, incidents and platform contracts.
Tools
Prefix Cache Auditor
Client-side diagnostic for unstable prefixes and schema drift.
Tools
LLM Cost Calculator
Estimate effective cost with cached input tokens.
Tools
AI Quality Gate Checklist
Local readiness review before rollout.
Templates
AI Scenario Migration RFC
A template for moving a scenario from MaaS to self-hosted or hybrid.
Templates
Planned
Model Release RFC
Release notes, quality gate, fallback and cost profile for model changes.
Templates
Planned
Eval Report
Dataset, error taxonomy, regression deltas and rollout recommendation.
Templates
Non-prod Cost Sheet
Production, stage, debug, canary, reserve and cost per accepted outcome.
Templates
Execution Boundary Matrix
A reusable matrix for MaaS, de-identified MaaS, self-hosted, batch and hybrid.
Templates
Planned
AI Incident Postmortem
A postmortem template for prompt, model, tool, cost and safety incidents.
Platform layers
The map has twelve responsibilities. If one is missing, a product team will implement it locally.
L01
Product use cases
Scenario intake, user value, risk profile, acceptance criteria.
L02
AI Gateway
Unified API layer for auth, quotas, routing, policy, cost attribution.
L03
Provider strategy
MaaS, OpenRouter-style research loops, self-hosted and hybrid decisions.
L04
Execution-path routing
Choosing the lane: small model, large model, RAG, agent, human review or policy denial.
L05
Inference runtime
LLM, STT, embeddings and reranker serving with latency and throughput budgets.
L06
Caching
Prompt cache, prefix cache, KV-cache and stable request shape.
L07
Model lifecycle
From research to shadow, canary, production, rollback and retire.
L08
Evals and Quality Gate
Datasets, eval suites, regression checks, canary rollout and feedback loops.
L09
Observability
Traces, tokens, TTFT, TPOT, fallback events, safety events and feedback.
L10
Economics / FinOps
Scenario cost, cached tokens, retries, GPU utilization and cost per accepted outcome.
L11
Guardrails / Security
Policies, PII, prompt injection, tool risks and audit trail.
L12
Operations / Ownership
Ownership, SLOs, incidents, capacity planning, runbooks and platform DevEx.
Start points
Foundational chapters for the first public version.
Production AI Platform Map
The responsibility-first map: from use cases to gateway, inference, evals and ownership.
AI Platform Maturity Model
A leadership framework from one API key to AI-native operating model.
MaaS vs Self-hosted
A strategy chapter for choosing managed, self-hosted or hybrid serving.
Semantic Router
How to choose direct, RAG, agentic or human-review execution.
Tools and templates
The handbook is meant to become maps, checklists, templates and small diagnostic tools.