Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

title: AI Engineering Handbook author: rudakow.wadim@gmail.com date: 2026-04-13 options: version: “3.1.0” birth: 2025-10-19 token_size: 3371

title: AI Engineering Handbook author: rudakow.wadim@gmail.com date: 2026-04-13 options: version: “3.1.0” birth: 2025-10-19 token_size: 3371

AI Engineering Handbook

This repository is a Documentation-as-Code hub for building production-grade AI systems. In the era of AI-backed development, AI systems — RAG pipelines, coding agents, AI assistants — consume documentation as their primary input. Poor documentation produces poor AI outputs. This repository treats documentation with the same rigor as production source code: versioned, tested, lifecycle-managed, and machine-readable.

Content is generated through a hybrid LLM methodology, cross-validated by multiple models, and reviewed by a human before promotion. Architectural decisions are governed by ADRs with machine-readable YAML frontmatter, enabling AI agents to filter and retrieve decisions by status, date, and scope.

Mission: To build reliable, machine-readable documentation infrastructure that AI systems can trust as authoritative input — while enabling MLOps engineers and AI architects to construct audit-ready AI systems using a hybrid LLM+SLM methodology.

What’s new?

v3.1.0 — “The Industrialization of Intelligence”

v3.0.0 — “The Agents Emerge”

v2.8.0 — “The Prompt Physics”

Live Documentation Site

The rendered knowledge base is published on GitHub Pages (ADR-26022). GitHub is the source of truth; the site is the rendered view for reading.

Documentation as Source Code for AI

Traditional documentation is written for humans to read after the fact. In AI-backed development, documentation is the primary input that AI systems consume to produce outputs. This makes documentation quality a first-class engineering concern: stale docs produce hallucinations, unstructured metadata makes retrieval blind, missing lifecycle allows noise to accumulate.

This repository implements the Documentation-as-Code paradigm:

This stands in direct opposition to pre-AI knowledge bases — collections of articles with no enforced structure, no traceability, and no automated verification.

See the full rationale in architecture/manifesto.md.

In this paradigm: prompts = source code, articles = build artifacts, reviews = QA gates.

Architectural Governance

ADRs are the main context for development in this repository. Every structural, methodological, or tooling decision is recorded as an ADR before implementation.

Hub-and-Spoke Ecosystem

This repository is the standards hub — it holds conventions, specifications, and ecosystem ADRs (ADR-26020). Implementation lives in independent spokes:

Methodology

Content generation follows a tool-agnostic, cognitive-model approach:

Coverage

Content is organized around five layers of AI systems. Security is a cross-cutting concern woven into all layers, not a separate layer.

  1. Execution & Optimization: CPU/GPU hybrid pipelines, VRAM/RAM management, NVIDIA tuning

  2. Model Development: SLM selection, tokenization, embeddings

  3. Prompt Engineering: Modular design, XML schemas, template lifecycle, consultant prompts

  4. Orchestration: RAG, workflow chaining, agent specialization, structured output

  5. Context Management: Vector stores, hybrid retrieval, indexing strategies

Read more in A Multi-Layered AI System Architecture.

Repository Structure

.
├── 0_intro/                # Introductory material, high-level overviews
├── ai_agents/              # Real-world agent framework analysis (assembled product)
│   ├── context_management/ # Context window strategy analysis
│   ├── session_history_management/ # Session history analysis
│   ├── skills/             # Skill system analysis
│   └── tooling/            # Tool architecture analysis
├── research/               # External source code repos for comparative study
│   ├── ai_coding_agents/   # Nested git repos (external agent source code)
│   └── ai_infrastructure/  # Nested git repos (non-agent AI projects)
├── ai_system_layers/       # Core content organized by layer (engine parts catalog)
│   ├── 1_execution/        # CPU/GPU optimization, CUDA, VRAM management
│   ├── 2_model/            # Model selection, tokenization, embeddings
│   ├── 3_prompts/          # Prompt-as-Infrastructure: formatting, versioning, consultants
│   ├── 4_orchestration/    # Workflow, agent specialization, structured output
│   └── 5_context/          # RAG, knowledge bases, hybrid retrieval
├── architecture/           # Architectural governance
│   ├── adr/                # Formal ADR documents (YAML frontmatter, machine-readable)
│   ├── evidence/           # Architecture Knowledge Base (analyses, sources, retrospectives)
│   ├── packages/           # Spoke package documentation and PoCs
│   └── post-mortems/       # Retrospectives on failures and lessons learned
├── misc/                   # Non-core work
│   ├── in_progress/        # Drafts, experiments, work-in-progress
│   ├── plan/               # Implementation plans (saved between context switches)
│   └── pr/                 # Telegram channel posts, review checklists
├── mlops/                  # MLOps tooling and pipelines
│   ├── ci_cd/              # GitHub Actions, quality-gate scripts
│   └── security/           # Secret scanning, compliance checks
├── security/               # Centralized security policy hub (placeholder)
└── tools/                  # Supporting utilities
    ├── configs/            # Tool configuration files
    ├── docs/               # Documentation generation helpers, style guides
    ├── scripts/            # Automation: link checking, ADR validation, Jupytext sync
    └── tests/              # Test suites for scripts

Toolchain & CI/CD

The repository uses automated validation to enforce documentation quality:

Authorship & Contact

My name is Vadim Rudakov, I am a Systems Engineer and AI Methodology Architect from Sverdlovsk oblast, Russia. This repository solves a specific problem: the fragility of AI prototypes and the risk of institutional “tribal knowledge.”

The repository is licensed under GPLv3 for its core assets and CC-BY-SA 4.0 for the article content.

Contact Information

Vadim Rudakov, rudakow.wadim@gmail.com

AI Engineering Book
LICENSE