AI/ML Engineer: An Expert's Guide to the Judgment Calls Behind Production ML, LLM Economics, and the Climb to Staff and Principal — Ebook (EPUB + PDF)
Past the mid-career mark, almost nothing you're handed comes with a clean answer attached. A recruiter quotes a specialization premium that scatters from 15% to 60% depending on who measured it. Your offline metrics improve and your online numbers get worse, and the team is looking at you to say which one is real. A system you're about to inherit has had three prior owners, and you have maybe two questions before you sign to find out if it will grow your career or quietly cap it. This book doesn't pretend any of that resolves into a single number. It teaches the judgment to make a defensible call anyway — reading a recruiter's percentage for what it's actually measured against, triaging a "hallucinating" RAG system by layer instead of guessing, and knowing when production results override the offline metric you're proud of.
It delivers that through the actual mechanics, not slogans: the retrieval-augmentation-generation triage order for debugging a customer-support bot that "makes things up," the hardware-and-inference-engine coupling decisions (H100 versus A100, vLLM versus TensorRT-LLM) that decide whether your LLM bill is defensible or three times what it should be, and the position-bias/exposure-bias/feedback-loop diagnosis for a re-ranking model whose online numbers won't agree with its offline ones. The closing chapters turn the same discipline on your own career: the two diligence questions — how much of the stack is proprietary, how old is the system and how many owners has it had — that tell you whether an offer builds transferable skill or quietly spends it down, and what actually changes once you're the one making the business case for a governance framework to a VP who's never heard of taint-tracking.
Inside
- The retrieval-augmentation-generation triage order for debugging a RAG chatbot that "hallucinates" — and why hybrid search, not a prompt tweak, is usually the actual fix
- The hardware-tier and inference-engine coupling decision (H100/A100/consumer GPU against vLLM/TGI/TensorRT-LLM), plus the caching-and-routing levers that move a production LLM bill by 40-70%
- The three mechanisms behind an offline-online metrics divergence — position bias, exposure bias, feedback loops — and how to run the rollback-versus-roll-forward call as a leadership decision instead of an argument
- A real defense architecture for agentic systems: capability-scoped tools, a deterministic policy layer that blocks an action before it fires, and defenses against memory poisoning, built around a real indirect-prompt-injection incident
- The Promotion-Politics Gap and the Proprietary-Tooling Trap — the two structural diligence questions that predict whether an offer will grow your career or cap it, before you sign
- The path from senior engineer through AI Solution Architect to Staff and Principal, told through architecture decision records, governance-framework ownership, and making a technical case land with someone who will never read it
Written for engineers who can already build a model, stand up a pipeline, and read a stack trace — the ones standing at the fork between the technical track and management, or close enough to it that the next decision is live rather than distant.
What you get: the complete ebook as EPUB (Kindle, Kobo, Apple Books and every reader app) and as PDF, delivered the moment you pay. Book 3 of the From Experience to Evidence AI/ML Engineer series.
Digital products are delivered the moment you pay. By completing checkout you ask for immediate delivery and give up the 14-day right of withdrawal that applies to consumers in the EU and UK; the refund policy at fromexperiencetoevidence.com/en/legal/refunds/ still applies.
