Building AI Products from Scratch— Volume 3: Testing, Performance, Observability, Security, Privacy, and Responsible AI - Tapa blanda

Sinopsis

Building AI Products from Scratch—Volume 3
Testing, Performance, Observability, Security, Privacy, and Responsible AI

Making an AI product work is only the beginning. Making it reliable, secure, observable, and trustworthy requires engineering discipline.

AI systems behave differently from conventional software. Model outputs can vary. Prompts can change behavior. Retrieval can return incomplete or conflicting information. Agent workflows can fail in unexpected ways. Performance and cost can change with usage. Security and privacy risks can emerge across models, data, tools, and integrations.

Building AI Products from Scratch—Volume 3 focuses on the engineering practices required to move AI products toward dependable production operation.

This volume provides a systematic approach to testing, measuring, observing, protecting, governing, and continuously improving AI systems.

• Design testing strategies specifically for AI products
• Build deterministic unit and integration tests around AI workflows
• Evaluate prompts, models, retrieval, and generated outputs
• Detect hallucinations, regressions, quality degradation, and behavioral changes
• Establish meaningful AI evaluation datasets and quality thresholds
• Engineer latency, throughput, token consumption, caching, and cost
• Design performance budgets and capacity strategies for AI workloads
• Build observability using logs, metrics, traces, and AI-specific telemetry
• Monitor model behavior, retrieval quality, failures, drift, and production changes
• Identify and defend against prompt injection, jailbreaks, data leakage, and AI abuse
• Secure APIs, tools, identities, secrets, model access, and AI workflows
• Design privacy-aware data flows, retention policies, consent mechanisms, and access controls
• Understand compliance considerations for AI systems and sensitive data
• Build responsible AI practices around fairness, transparency, explainability, human oversight, and risk
• Establish production quality gates and continuous monitoring processes

Traditional software testing often assumes deterministic behavior. AI systems require a broader definition of quality.

A system can be available but produce poor answers. It can pass functional tests while becoming more expensive. A model can perform well on one evaluation dataset while failing on another. A retrieval system can return relevant information while violating access boundaries.

This volume therefore treats AI quality as a multidimensional engineering problem.

Testing provides evidence.

Performance establishes operational boundaries.

Observability reveals what is happening.

Security controls what can go wrong.

Privacy protects valuable information.

Responsible AI addresses system impact and accountability.

The goal is not to assume that AI will behave perfectly. The goal is to build systems where important failures can be detected, investigated, controlled, and improved.

This volume is written for AI/ML engineers, software engineers, full-stack developers, system architects, DevOps and SRE teams, security professionals, privacy and compliance teams, engineering managers, product leaders, founders, and developers moving AI systems into production.

If you want to move beyond making AI applications work and learn how to make them measurable, reliable, secure, observable, privacy-conscious, and responsible, Volume 3 provides the engineering discipline required for the next stage.

Test with discipline. Measure with evidence. Observe continuously. Secure deliberately. Protect privacy. Build AI systems that remain accountable.

"Sinopsis" puede pertenecer a otra edición de este libro.