GuideIntermediate

Client-Facing AI Engineering Guide

More than half of AI engineer job postings ask for something that is not code: sitting with a client, finding out what problem they really have, agreeing on what "it works" means, explaining costs and risks to people who don't code, showing a prototype with its limits, measuring a pilot, and leaving the system in the hands of their team. In forward deployed and applied AI roles it is almost the whole job. This guide teaches that craft from the engineer's seat, with documents and code you can show in an interview. You learn it as an AI engineer at **Brújula Labs**, an applied AI consultancy, on an engagement with **Refacciones Ámbar**, an auto parts distributor with 14 branches in western Mexico. Its sales director asks for "a chatbot that answers repair shops 24 hours a day". Through discovery you find that the problem is a different one and turn it into a quote copilot for the parts counter: you observe the work, measure the baseline, scope the use case and its level of autonomy, agree on success as an acceptance set with thresholds the client signs, write a proposal with three options and their cost, build the prototype and demo it, run a pilot in shadow mode, deliver bad news on time, and hand off: operating manual, training, adoption tracking, and feedback for the product team. All the code is Python and runs on your laptop without an API key, with case data generated from a seed: you see the same numbers as the guide. The prototype reads messages with rules and uses `gpt-6-luna` only to read what the rules can't; the model never picks the part. The analogy that runs through the guide is a made-to-measure tailor: measure before cutting, baste before sewing, and deliver the suit with its care label.