Skip to content

Build AI Apps

An AI Assistant, built one capability at a time. Every lesson is labelled with who provides the capability, because the most expensive assumption a reader can make is that the framework ships the part they are about to write: xyberos.ai is a provider-neutral facade, a closed JSON Schema subset, an allowlisted tool layer, and a deterministic fake provider. There is no memory store, no document loader, no vector index, no intent classifier, and no agent loop in the package, and the lessons that need one say so and build it.

Each lesson runs with no model service and no network: the code you read is the code the suite runs, against FakeModelProvider, and the output you see is the output it really produced. A live model is an opt-in appendix rather than a lesson.

Lessons

# Lesson You will learn Provided by
1 An application that is AI-ready A model behind a seam, and an app that runs without one Built-in
2 What a provider declares Capabilities, checked before the call rather than after Built-in
3 The first call Messages, responses, model names, and token usage Built-in
4 Switch providers without touching the application Why the facade exists, and what it costs Built-in
5 Structured output A closed schema subset, and three distinct refusals Built-in
6 Failures, timeouts, and retries The error taxonomy, and which failures are transient Built-in
7 Conversation history Context the caller assembles, and the window it sends You build it
8 Memory A table, a query, and a decision about the next request You build it
9 Knowledge Loading documents, and where a passage is cut You build it
10 Retrieval over embed A similarity query you build, and why it must be measured You build it (over embed)
11 Tools An allowlist, arguments checked against a schema, and events Built-in
12 Intents Classification with structured, and the rules that come first You build it (over structured)
13 Workflows Explicit steps, branches, deadlines, and a renderable graph Built-in
14 Bounded agent behaviour Tools plus a window, a budget, and the escalation path You build it (tools + workflow)
15 Testing the AI path Every prompt, tool, and failure with no model service Built-in (FakeModelProvider)
16 Evaluating quality Accuracy, grounding, cost, and latency as rows You build it
17 Securing AI applications Prompt injection, tool permissions, and redaction Built-in boundaries, application code

All seventeen lessons are written. The roadmap is generated from each page's own status, so it is the authority on what has landed rather than this table.

What you need

The core package, and nothing else: the setup from fundamentals lesson 1. To point any lesson at a real local model, add the one adapter that ships:

python -m pip install "xyberos[ollama]"

python -m pip install 'xyberos[ollama]'

Two things to know before it starts

Preview means preview. xyberos.ai is preview tier: the contract is settled and the surface may still change in a minor release. Every page repeats the tier of what it teaches.

The fake provider is not a stub. FakeModelProvider implements the provider contract, records the requests it received, and answers deterministically from a script. That is why the AI path is testable in CI, and lesson 15 is about using it rather than about avoiding it.