The AI-FDE Transition Path
From the client's site to AI delivery: a learning path for implementation, pre-sales, delivery and operations people moving into the AI-FDE role.
13 articles
The AI-FDE Transition Path · Introduction
The AI-FDE Transition Path · AI Engineering
First, Understand Exactly Where the Model Will Always Fail
In AI delivery, the first thing to learn is not how to tune a model, but what it can and cannot do and where it always fails. This piece lays out four failure classes and what 'knowing it' looks like.
Prompts Aren't Magic — They're Code You Can Regression-Test
Prompts are the one place in an AI project that anyone can change and nobody owns. Treat them like code: keep them in files, log every change, and run a regression set.
Context Engineering: What the Model Can Actually See
A model has no memory: every turn it sees only the text you reassemble and send. Understand what context contains, how it is capped, compressed and billed — and you know what to put in and what to leave out.
When the Knowledge Base Is Wrong, the Model Usually Isn't
RAG is an open-book exam for the model. Most of the work happens on the retrieval side, so when answers are wrong, split the pipeline, print the intermediate results and use a triage table — don't start by swapping the model.
80% of RAG Problems Start in Document Parsing
The dirtiest, most tedious and highest-ROI part of RAG: four killers — scans, tables, multi-column layouts and heading hierarchy — plus chunking, a health-check sheet and a Dify implementation map.
Tool Descriptions Are Manuals Written for the Model
What the model lacks isn't intelligence but data, arithmetic and hands. How to classify and connect tools — and why one line of tool description decides whether it gets called.
How a Model Actually Uses a Tool
The model never calls a tool. It only emits text saying what it wants to call and with which arguments. A four-round request walkthrough shows execution, result injection, cost and where MCP actually fits.
A Python Primer: Get the Environment Right First
For FDEs who don't write code: versions, virtual environments, offline package installs, reading tracebacks, understanding unfamiliar code in five minutes, and keeping acceptance authority when AI edits code.
Turning an Internal System into an MCP Server
From roles and the boundary between tool calling and MCP, to whether wrapping is worth it, four acceptance gates, and twenty-odd lines of working code.
Designing an Agent from Scratch (Part 1)
An agent is not a smarter model but a program that lets the model act: one loop, three layers of brakes, a minimal loop, stop lines, observability, and four file tools.
Designing an Agent from Scratch (Part 2)
Four layers of shell interception, retrieval and context trimming, slash commands, and sub-agent permissions: making an agent controllable, not just runnable.
Designing an Agent from Scratch (Part 3)
Three extension paths (plugins, MCP, Skills), program-level bookkeeping and pre-answer checks, per-call logs and long-term memory, plus an honest acceptance checklist.