Books

RAG, Agent Bricks, the Multi-Agent Supervisor with MCP, Lakebase, MLflow 3, Lakehouse Monitoring, Feature Store, Vector Search. Every AI surface Databricks shipped at GA in 2025 and 2026, taught by a practitioner, current to 2026. What you will learn - Build RAG pipelines with Vector Search, embedding models, and citation grounding- Ship Agent Bricks for classification and information extraction- Orchestrate specialist agents with the Multi-Agent Supervisor and MCP- Use Lakebase as the operational Postgres layer for AI apps and agents- Detect data and model drift with Lakehouse Monitoring; wire alerts to retraining- Manage the ML lifecycle with MLflow 3 and the UC Model Registry- Govern features across training and serving with Feature Store (offline + online)- Serve foundation and custom models with AI Gateway controls Who this book is for Data engineers, ML engineers, and AI/ML architects who know PySpark and the Databricks platform and now need to ship production AI. Volume 3 is the recommended prerequisite. Table of Contents 1. Databricks SQL in Production. Warehouses, materialized views, three latency signals (admission, compilation, execution), the full dashboard backend wiring.2. External BI: Tableau, Power BI, dbt. Performance tips that take a dashboard from sluggish to instant, dbt configuration at incremental scale, the seam between BI and the lakehouse.3. AI/BI Dashboards. Anatomy of a Lakeview dashboard, draft vs published flow, the Dashboard Agent's reliable patterns, the five-grant permission model.4. Genie: Natural-Language Analytics. Grounding sources, the priority rule, the SQL Genie actually writes, the questions Genie answers cleanly versus the ones that confuse it.5. AI SQL Functions. ai_query, ai_parse_document, ai_extract for PDFs and HTML, univariate forecasts, the daily cost math for production AI SQL pipelines.6. Model Serving. Endpoints, the three fields that decide capacity and cost, the chat-completion payload, the five moving pieces of a production recommender.7. Foundation Models. Five major providers, the External Models config, the vendor-swap pattern (Claude to Gemini in hours, not weeks), the three habits that keep swap cost low.8. Vector Search and RAG. Six delta-sync arguments, three chunking strategies compared, the RAG function your app imports, end-to-end answer evaluation with traces.9. MLflow 3 and UC Model Registry. Versions, aliases, tags (and what each is not for), five tracking calls and what each one writes, the experiment-to-production lifecycle.10. Feature Store. Why SDP is the right producer, the six-file project layout, four parity-failure classes between offline and online stores and what causes each.11. MLOps as a Practice. Seven sources every incident reads from, three deploy patterns (canary, shadow, blue-green), three retrain strategies, five golden signals for an ML endpoint.12. Lakehouse Monitoring: Drift Detection. Six monitor parameters, the loop from drift alert to retraining, what to do when the baseline table is missing.13. Distributed Deep Learning. Three signals that force distributed training, picking the flavor (data, model, hybrid) from the bottleneck, four pieces of GPU memory worked out for a 7B model.14. Agent Bricks. Declarative classification and information-extraction agents, eval-set ingredients, the pre-compute pattern that makes small seed sets work.15. Multi-Agent Supervisor and MCP. The supervisor build, synthetic-turn evaluation, three real conversations end to end, the auth-passthrough chain across child agents.16. Lakebase: Operational Postgres for AI. Five alternatives compared, sub-10ms reads for AI apps, the lineage from Delta source through SDP into Postgres and onward to the endpoint.17. Capstone: Retail Intelligence App. Ten stages, each anchored to an earlier chapter. The smoke test that confirms every stage of the platform is reachable, the new-data path through the recommender.18. Certification and What's Next. The certification paths that actually map to the book, and the reading list the on-call team uses when something breaks.

AWS replaced Amazon Q with Kiro — an AI IDE that writes specs before code. 14 chapters covering spec-driven development, AI agents, hooks, MCP integrations, AWS deployment, and the honest comparison with Cursor and Copilot. The definitive guide.

The tech job market changed. Entry-level dropped 73%, AI screens resumes, interviews test system design. This is the no-BS guide: portfolio, resume, LinkedIn, interviews, negotiation, and your first 90 days. Written by someone who hires engineers.

Kubernetes exposes plenty of state, but operators still have to reconstruct operational meaning from scattered Pods, Services, EndpointSlices, Events, PVCs, owner references, and status fields. This ebook uses `kctx`, a small read-only Kubernetes context engine, to show how deterministic entities, relations, signals, graphs, namespace snapshots, CRD adapters, and stable JSON contracts can turn raw cluster data into reusable context for humans, tools, and AI agents. It is written for SREs, platform engineers, Kubernetes operators, infrastructure developers, and AI tooling builders who want better primitives than raw YAML and improvised troubleshooting pipelines.

Stop maintaining CLAUDE.md and AGENTS.md as two files that drift apart. This handbook documents which file each of nine tools reads — checked on 2026-06-02 against each tool's official docs — and ranks six ways to keep one instruction file in sync, from an @AGENTS.md import to a symlink to a CI drift check. It is also precise about where official support still is not: issue #6235 was closed as completed, but the official CHANGELOG does not mention AGENTS.md anywhere, and on v2.1.246 AGENTS.md is still not loaded into the model's context the way CLAUDE.md is.

Between March and April 2026, Claude Code subscribers lived through six weeks of stacked regressions: cache TTL shortened from sixty minutes to five, a tokenizer that charged 1.35 to 1.46 times more for the same prompts, Pro access that briefly disappeared, and a third-party tool ban. This playbook turns that window into a decision instead of a guess: five measurable triggers, three paths — stay and fortify, switch, or build your own stack — and a decision tree that returns one specific recommendation, with a 30-day cost projection for each path. Edition 2 is 251 pages: ten chapters, six path supplements and four appendices, reconstructed from public GitHub issues, Hacker News threads, and independent measurement posts.

Once the five essential hooks are in place, the incidents that keep happening are the guards themselves failing, silently. Nine chapters collect the hooks (Git, credentials, token spikes, autonomous operation, multi-agent work), a chapter on hooks that do not fire, eight dated incidents from the author's own unattended operation, and a 50-point audit you can run against your own setup.

92% of developers use AI daily. Entry-level jobs dropped 73%. This isn't a story about replacement — it's about transformation. Data-driven guide with a 30-day action plan for developers who want to thrive, not just survive.

Your AI prototype works. Now ship it. Most AI frameworks are built for exploration. kdeps is built for production. Define your agent in YAML, declare its dependencies, and deploy it anywhere — Docker, Kubernetes, a standalone binary, an edge device — without rewriting a line when you switch LLM providers. AI Appliances is the hands-on guide to building autonomous AI agents and multi-agent systems with kdeps: deterministic pipelines, real error handling, real deployment, and no vendor lock-in. Write YAML. Run anywhere. Own everything.

Build defensible evaluation systems for Gen AI, RAG, agents and multimodal applications using datasets, metrics, release gates, monitoring and seven interactive labs.

A leader at a company opens Slack one morning. The brief their virtual employee wrote looks normal: same headers, same prose voice, same cadence as every other brief that week. The last line ends mid-sentence. The API returned `stop_reason: "max_tokens"`. The system shipped it anyway. No exception. No log line. No retry. That bug doesn't look like the normal kind. This book is about why, and what to build instead.

45% do código gerado por IA tem vulnerabilidades. Este livro cobre prompt injection, vazamento de dados, codificação segura com ferramentas de IA, hardening de servidores MCP, sandboxing de agentes e compliance OWASP LLM. 14 capítulos com demonstrações de ataque e código de defesa.

Guia prático para desenvolvimento agêntico com Claude Code, MCP, Claude API e Agent SDK. De modelos mentais a arquiteturas de produção — tudo que um dev experiente precisa para trabalhar de forma agêntica, não apenas usar ferramentas de IA.

A hands-on guide to building with Claude Code, MCP, the Claude API, and the Agent SDK. From mental models to production architectures — everything an experienced developer needs to work agentically, not just use AI tools.