Becoming a Forward Deployed Engineer

The role: part engineer, part operator, part account owner

What an FDE actually does — embedded inside a customer’s org to make an AI product work in a real business, not just demo it. Why the role exists, what separates it from a solutions engineer or consultant, and the compensation that comes with it ($205K–$486K, staff clearing $630K+).

Why every AI lab copied Palantir in the same week

Anthropic ($1.5B) and OpenAI ($10B) both launched deployment ventures within days of each other. The template is Palantir’s FDE model. Understanding why the labs are betting on this go-to-market tells you the role is structural, not a trend — and where the hiring is heading.

The actual path from where you are now

The honest route from SWE, data, or consulting into an FDE seat. What transfers, what you have to build, the 60–120 day prep timeline, and the Palantir-style case interview that 60% of candidates fail. Written for someone who isn’t already at an AI company.

The 5 portfolio projects that signal readiness

Interviews want proof you can build against a client’s messy data and ambiguous problem. These five projects show exactly that — what to build, what to include, and what NOT to do. The differentiator most candidates miss.

The core deliverable: one routing layer for all the data

An MCP server is named explicitly in the Anthropic FDE job listing. This is the thing you build on day one at a client: a single interface that connects Claude to their CRM, billing, support, and comms — with shaped responses and access control baked in.

Connecting agents to real production systems

Moving from a notebook demo to an agent that touches a client’s live systems. The patterns for safe writes, auth, and the failure handling that separates a pilot from something that survives contact with production.

Access control: the enterprise non-negotiable

When you wire AI into a client’s data, "everyone gets everything" gets you fired. Tool-level permissions enforced in middleware — not by the model — are the only reliable solution, and the thing every enterprise security review will probe. How to build it right.

The full deliverable, end to end

What the system you hand off actually looks like: the five-layer stack, the target metrics, and what separates a deployment that lasts after you leave from one that quietly dies in month three. The architecture an FDE is ultimately accountable for.

Scoping: the first week decides the next three months

Every engagement that goes wrong was scoped wrong in week one. The five-day discovery structure, the three queries to run on day one, how to get a measurable definition of "working" out of the person who controls the budget, and the sizing multiplier that is actually real.

When the data is worse than they said

Not an edge case — the default. Six checks that turn "the data is messy" into numbers you can act on, the framing that delivers the bad news without detonating the engagement, and what you can still ship when a third of a critical field is null.

The handoff that survives your departure

The measure of the work is what still runs in month six, after a model deprecation, a schema change, and the departure of the one person who understood it. The four things that actually kill delivered systems, and the symptom-first runbook that prevents three of them.