What is From Model to Product about?
Short answer
From Model to Product: Foundations for Practical AI Engineering, by Dr. Dulan Dias, is a 403-page book about turning AI models into dependable products. Its argument is that the hard part of AI is not the model but everything around it: problem framing, data, APIs, storage, deployment, monitoring, reliability, cost, teams and governance. The book covers back-end and production engineering; it does not cover front-end work or agentic AI.
By Dr. Dulan Dias, co-founder and CTO of Tellme AI
Key facts
- From Model to Product was published independently in March 2026: the hardcover on 16 March and the paperback on 21 March.
- From Model to Product runs to 403 pages and 18 chapters, arranged in five parts.
- The paperback ISBN of From Model to Product is 979-8-2519-9095-9 and the hardcover ISBN is 979-8-2519-9795-8.
- The preface limits the book to back-end and production engineering for AI products; front-end frameworks, UI implementation and client-side design are outside its scope.
- Part III (Chapters 9 to 12) builds one running product, HomeSort AI, a mobile-first web app that tells UK households how to dispose of an item from a camera photo.
- In the book's running example the backend decides and the language model explains: rules and a classifier choose the disposal category, and the model only answers the user's questions about it.
- The book's Wikidata identifier is Q141865687.
Definitions
- Production AI
- In the preface's terms, not a model endpoint that returns plausible output but an end-to-end operating system for decisions, data and services: stable pipelines, robust service contracts, reproducible deployment, observability, incident response and governance.
- HomeSort AI
- The book's running example from Chapter 9 onwards: a mobile-first web app that recognises a household item on the device, applies regional disposal rules stored in a database, falls back to a server model when unsure, and uses a language model only to explain.
- Decision hierarchy
- The book's pattern for answering each request with the simplest method that can: rules first, then a classifier, then a fallback model, then a language model, and finally a safe response.
What is the book's argument?
The book argues that building a model is now the easy part, and that most AI projects fail in the gap between demo and deployment, when noisy data, uneven traffic, unclear ownership and rising cost meet a system designed around a model rather than a product. The preface puts it plainly: “building something that works once is easy. Building something that keeps working is the real challenge.”
Its answer is a way of thinking rather than a stack. Treat the model as one capability inside a product, start from the problem and its constraints, and use the simplest level of intelligence that meets the requirements, moving from rules to classical machine learning, deep learning and language models only when the evidence justifies it.
That judgement is then tested on one product, built across Part III and operated through Part IV, so the advice is shown working on the same system rather than in isolated snippets.
How is the book organised?
The 18 chapters sit in five parts, moving from judgement to architecture, then to a worked build, production and the people around it.
| Part | Chapters | What it covers |
|---|---|---|
| I. Thinking Like an AI Engineer | 1 to 3 | The product mindset, problem framing and choosing the level of intelligence a problem needs. |
| II. Architecture & Design Fundamentals | 4 to 8 | Data, model, API, system and storage architecture. |
| III. Building Useful AI Solutions | 9 to 12 | One product built end to end: edge vision, backend logic, a grounded language model and the whole architecture. |
| IV. Deployment & Productionization | 13 to 15 | Packaging and release, monitoring, reliability, scaling and cost. |
| V. People, Process & Professional Practice | 16 to 18 | Teams, ethics and governance, and the principles that outlast today's tools. |
What does each chapter cover?
| Chapter | In one line |
|---|---|
| 1. The AI Product Mindset | Why an impressive model is not a product, how production AI differs from research and demos, and the habits that close the gap. |
| 2. Problem Framing and Requirements | Turning a vague business request into a measurable problem with requirements, assumptions and a check for non-AI alternatives. |
| 3. Choosing the Right Approach | Choosing between rules, search, classical ML, deep learning, language models and hybrids, using a complexity ladder and the accuracy, latency and cost triangle. |
| 4. Data Architecture for AI | How data moves from sources to model inputs, and how training-serving skew, leakage, versioning and drift break that chain. |
| 5. Model Architecture Patterns | Structuring the model under real constraints, from model families to prompting, fine-tuning, RAG, optimisation and ensembles. |
| 6. API-Driven AI | Serving a model through a production API with FastAPI and Pydantic: schemas, authentication, async jobs, caching, rate limits, errors, versioning and health checks. |
| 7. System Architecture Patterns | Monolith first, request-response versus events, online versus offline inference, scaling, statelessness, and containers versus serverless. |
| 8. Storage and Databases for AI Systems | Choosing storage by workload: SQL versus NoSQL, vector indexes, caches, feature stores, storage tiers, replication and tested backups. |
| 9. Vision on the Edge | Starting HomeSort AI: an image classifier built from real datasets, run in the browser, with a confidence threshold and an explicit unknown class. |
| 10. From Prediction to Product Logic | The backend that turns a label and a confidence score into a rule-based decision, with sessions, stored rules and API keys kept off the client. |
| 11. Language Models in Real Systems | When to call a language model at all, and how to keep it grounded, consistent with the system's decisions, affordable and available. |
| 12. Building a Complete AI Product | Assembling the whole system, walking its pipeline step by step, and redesigning it as users, cost, latency or the problem change. |
| 13. Packaging & Deploying AI Systems | Reproducible environments, containers, CI/CD that versions code, models, prompts and rules together, and blue-green, canary and shadow releases. |
| 14. Observability, Monitoring & Model Health | Seeing whether a running system still works: layered monitoring, drift signals, dashboards, alerts and a cost spike case study. |
| 15. Reliability, Scalability & Cost Management | Fallback chains, retries, circuit breakers, graceful degradation, ways to scale inference, cost levers and load testing. |
| 16. Working in Teams | The seven roles on an AI product team, how shared words mean different things to each, managing ambiguity, and deciding when a system is good enough to ship. |
| 17. Ethics, Governance & Risk | Bias, fairness and explainability as engineering concerns, plus model versioning, registries, approval workflows and risk tiers. |
| 18. The Future of AI Engineering | What changes and what stays constant, the shift from model-centric to system-centric work, and how to keep learning as tools change. |
Who is the book for, and who is it not for?
The preface names software engineers, AI practitioners, data scientists, technical founders, product managers and self-taught developers who are moving from experiments to dependable products. It expects two kinds of reader: people with strong software backgrounds who want deeper AI judgement, and people who understand model behaviour but want stronger architecture and operations skills.
- A good fit if you have a prototype, notebook or AI feature that works in a demo and must now survive real traffic, messy data, failures and cost.
- A good fit if you built something with an AI coding assistant and need it to work for real users; the guide to taking a vibe-coded app to production applies the book to that case.
- A poor fit if you want the mathematics of model training or the internals of transformers.
- A poor fit if you want to build agents with tools and planning; the book does not cover agentic AI.
- A poor fit if you need front-end or UI guidance.
What will you be able to do after reading it?
The preface suggests one habit for getting value from it: read a chapter, map its ideas to your current system, then write down two decisions you need to make this quarter.
- Frame a business problem and its success measures before choosing a model or tool (Chapter 2).
- Decide between rules, classical ML, deep learning and an LLM for each decision in a product (Chapter 3).
- Design APIs, data pipelines and storage that hold up under concurrency and change (Chapters 4 and 6 to 8).
- Package, release and roll back an AI system, and monitor it for drift, fallback creep and cost growth (Chapters 13 and 14).
- Keep a system working when dependencies fail and costs rise, using fallbacks, circuit breakers and cost levers (Chapter 15).
- Work across engineering, product and operations, and apply versioning, approval and audit controls scaled to the system's risk (Chapters 16 and 17).
What does the book leave out?
Knowing the limits saves you buying the wrong book. From Model to Product does not cover:
- Front-end frameworks, UI implementation or client-side design, which the preface puts out of scope.
- Agentic AI: agents, tool use and planning loops. For those, start with what agentic AI is.
- Specific laws and regulations. Chapter 17 treats ethics, governance and risk as engineering practice but does not walk through any law.
- Deep learning theory. Models are trained and used through transfer learning and APIs; the mathematics of training is not taught.
Formats, ISBNs and where to buy
Both editions are independently published. As of November 2026 the book is sold on Amazon UK and Amazon.com, and the paperback is also sold directly from the book page for delivery to UK addresses.
| Format | Published | ISBN | Pages |
|---|---|---|---|
| Hardcover | 16 March 2026 | 979-8-2519-9795-8 | 403 |
| Paperback | 21 March 2026 | 979-8-2519-9095-9 | 403 |
Frequently asked questions
Is From Model to Product suitable for beginners?
From Model to Product suits readers who can already program and have built or used a model, rather than people new to programming. It explains ideas such as drift, fallbacks and confidence thresholds as it goes, and the preface says you do not need mastery of every topic on day one. Someone new to machine learning will get more from it after a hands-on introduction.
Does From Model to Product cover large language models and RAG?
From Model to Product covers language models as one component of a product. Chapter 5 compares prompting, fine-tuning, retrieval-augmented generation and API versus local models; Chapter 8 explains vector indexes; and Chapter 11 builds a grounded prompt, uses simple retrieval before generation without a vector database, and adds selective model calls, provider fallback and model routing.
Does From Model to Product cover AI agents?
From Model to Product does not cover agentic AI. It has no treatment of agents, tool use or planning loops; its language model chapters keep the model in an explaining role while rules and code make the decisions. Readers who want agents need another source for them, though the book's chapters on reliability, monitoring and cost still apply to any system that calls a model.
Which programming languages and tools does the book use?
From Model to Product uses Python for its examples. Chapter 6 serves models with FastAPI and Pydantic, Chapter 9 trains a MobileNet image classifier with TensorFlow and Keras, Chapter 10 keeps rules and sessions in MongoDB, and Chapter 13 packages the system with Docker and Docker Compose. Chapter 6 also explains why another team might sensibly serve models from Go, Java, Node.js or Rust instead.
Do you need to read the chapters of From Model to Product in order?
From Model to Product can be read either way. The preface says you can read in sequence for a complete system view, or jump straight to a problem area such as architecture, deployment, monitoring, data workflows or reliability. Part III is the exception: Chapters 9 to 12 build one product step by step, so read them in order.
Is there a Kindle or ebook edition of From Model to Product?
As of November 2026, From Model to Product is published in paperback and hardcover only, and the Amazon UK and Amazon.com listings show no Kindle edition. Both print editions run to 403 pages. The paperback can also be bought directly from the book page on this site for delivery to UK addresses.
Who wrote From Model to Product?
From Model to Product was written by Dr. Dulan Dias, co-founder and CTO of Tellme AI in Belfast, founder of EchonLabs and holder of a PhD in Artificial Intelligence and Data Science from Queen's University Belfast. The preface describes close to ten years of building software and AI systems behind it. Who Dr. Dulan Dias is gives the full record.
Sources
About the author
Dr. Dulan Dias is co-founder and Chief Technology Officer of Tellme AI in Belfast; founder of EchonLabs, grown from 2017 to more than 60 people before its acquisition in 2026; holder of a PhD in AI and data science from Queen’s University Belfast; and the author of From Model to Product.