Most AI tooling is built for the demo, not the shift. You get a working prototype, someone wants to put it in production, and suddenly you're asking: how do I validate inputs? How do
I retry a flaky API? How do I deploy this without rewriting it? How do I switch models without touching every file?
AI Appliances answers all of those questions with a single framework: kdeps.
kdeps lets you define AI agents as YAML workflows — deterministic pipelines of resources (LLM calls, SQL queries, HTTP requests, Python scripts, browser automation) connected in a
dependency graph. You declare what each step does and what it depends on. kdeps handles execution, validation, retries, error handling, and deployment.
This book is the complete guide to building with kdeps, from a first working agent to a multi-agent system running in production on Kubernetes.
You will learn how to:
- Build deterministic AI pipelines using a directed acyclic graph (DAG) of typed resources
- Wire LLM calls, SQL queries, HTTP clients, Python scripts, web scrapers, and browser automation into a single workflow
- Validate inputs, handle errors gracefully, and make agents resilient to real-world conditions
- Compose reusable components and build multi-agent agencies where specialized agents collaborate autonomously
- Deploy to Docker, Kubernetes, standalone binaries, or edge devices — from the same YAML, without changing a line
- Run local open-source LLMs via Ollama, or route to OpenAI, Anthropic, Gemini, Perplexity, or any compatible provider
- Debug with kdeps validate, kdeps doctor, hot-reload development mode, and trace logging
- Build native chat bots for Telegram, Discord, Slack, and WhatsApp — or process files from the command line — with the same resource DAG, just a different input source
What this book is not: a machine learning theory text, a prompt engineering tutorial, or a guide to any proprietary AI platform. It is a book about running AI in production as reliably as you run a database.
What you need: familiarity with YAML and basic command-line comfort. No prior AI/ML experience required.
