1. Forward-deployed LLM projects

A customer tells you they want "AI for our support team." There is no spec, no metric, and nobody has said which tickets, which users, or where the data may go. That is the daily work of a forward-deployed engineer (FDE): an engineer embedded with one customer who owns delivery end to end, a role Palantir made famous.

This tutorial teaches the delivery knowledge around the model: scoping the use case with a success metric, measuring a baseline and estimating return, testing a proof of concept on the customer's own data, fitting their security and deployment rules, and measuring impact after you hand it over.

This approach is useful when the request is vague and the customer's data and rules determine the design. If the measured baseline shows the savings will not cover the build, the right proposal is a smaller pilot or no LLM project at all.

In FDE and customer-facing AI engineering loops, this is often a whole case round. In platform roles it shows up as a few follow-ups. A mid-level answer narrows the ask, measures a baseline, and judges the prototype on a sampled evaluation set. A senior answer also plans data access and deployment around residency rules, starts the security review early, and measures impact against a comparison group. Prompting, retrieval, and scoring methods live in their own tutorials, and sales and contracts are out of scope here.

We'll use one engagement throughout. A mid-size project-management software-as-a-service (SaaS) company has 40 support agents handling about 12k tickets a month, with a median handle time of 11 minutes and a loaded agent cost of about $45 an hour. The ticket mix is how-to 40%, billing 20%, bug reports 25%, and refunds and account changes 15%, and every ticket lives in a helpdesk and knowledge base inside the customer's cloud account in the EU.