Artificial Intelligence
Forward Deployed Engineer / Applied AI Engineer
The biggest cluster in the entire dataset, and the one with the least prep material written for it.
What the job actually is
You are embedded with a customer, and you write the code that makes a general model work inside their specific mess — their schemas, their auth, their compliance boundary, their idea of what "correct" means. Roughly half the work is engineering and half is consulting: figuring out what the customer actually needs versus what they asked for. The role exists because frontier models are general and enterprises are not, and that gap does not close by writing better documentation.
Titles this hides behind
Who is hiring it
OpenAI, Anthropic, Palantir, Sierra, Harvey, Glean, Scale AI, Writer, Instabase, ElevenLabs. Frequently split by vertical — financial services, public sector, healthcare and life sciences, retail, manufacturing, media and games — which is a useful signal: pick the vertical where you already speak the language. In India the same shape appears at Sarvam, Turing and Observe.AI.
How to recognise it in a JD
Look for customer-facing, onsite, travel,
stakeholder, engagement, ambiguity, and
prototype to production. If the requirements list names a vertical
industry before it names a language, you are reading an FDE posting regardless of
what the title says.
The loop
Five to eight stages across three to six weeks. At OpenAI and Palantir, roughly half the total evaluation weight sits on non-coding rounds.
Prepare like this
- Build one agent against genuinely messy data — not a clean demo dataset. Inconsistent schemas, missing fields, three date formats. The stories that land in the case study round come from having actually hit this.
- Practise scoping out loud, on a timer. Take a one-line business problem, spend 45 minutes talking through decomposition to a recording. Listen back. Most candidates fail this round by solving too early rather than by solving wrong.
- Prepare three "they asked for X, the real problem was Y" stories with the discovery moment in each, and what it cost to have found it late.
- Be able to whiteboard an enterprise integration: where the data lives, how auth flows, what happens to PII, what the latency budget is, who pays for tokens, and what runs inside their VPC.
- Know the failure modes cold — retrieval returning plausible-wrong context, tool calls failing silently, cost blowing up on retries.
Be ready for
- "A customer says the model is wrong 30% of the time. What do you do first?" They want you to ask what "wrong" means and how it was measured, before proposing anything.
- "You have two weeks and an executive demo. What do you cut?" Testing whether you scope to a demonstrable slice or promise everything.
- "The customer's data is far worse than they told us. Walk me through the conversation." Testing whether you can deliver bad news without either sugar-coating or blaming.
- "How would you know this deployment succeeded, six months from now?" Testing whether you think in outcomes or in shipped features.
Red flags in the posting
Pre-sales in disguise. If the JD never mentions a codebase, code review, or shipping, this is a solutions consultant with an engineering title and an engineering interview but not an engineering career path.
Unbounded travel. "Travel as needed" with no percentage. Ask for the number in the recruiter screen; a real answer is 20–40%, and evasion is informative.
No engineering manager in the loop. If every interviewer is from sales or customer success, your performance reviews will be too.
Compensation
A day that is actually representative
| Time | What you are actually doing |
|---|---|
| 09:00 | Customer standup. Not your team's — theirs. You are embedded, and the first fifteen minutes tell you what changed overnight in their world. |
| 10:00 | Build. Writing the connector that gets their data into a shape the model can use, which is most of the actual engineering and none of the demo. |
| 13:00 | A meeting you did not expect, with someone senior who has just heard about the project and has an opinion. Handling this well is the job. |
| 15:00 | Reality check. What you promised for Friday is not going to be ready, and you decide now whether to cut scope or move the date — and tell them today rather than Thursday. |
| 16:30 | Write up. A short note to your own company on what this customer needs that the product does not do. This is how FDE work turns into roadmap, and it is the part that gets you promoted. |
Real cases worth being able to discuss
The original version of this role: engineers embedded with customers, building on a platform rather than delivering a fixed product. OpenAI, Anthropic, Sierra and others adopted it almost wholesale because frontier models have the same gap between general capability and specific environment.
Use it to say: “The reason the model exists is that the distance between a capable model and a working deployment is mostly other people's data and constraints, and you cannot document your way across it.”
A tribunal held the airline liable for its chatbot's incorrect advice; the argument that the chatbot was a separate entity failed.
Use it to say: “I raise it early with customers now, because it changes the conversation about guardrails from a technical preference to a liability question they already care about.”
A dealership assistant was talked into agreeing to sell a car for one dollar and calling it binding.
Use it to say: “It is the cheapest way I know to explain why the control belongs on what the system may commit to, not on the prompt.”
A three-week plan
Week 1 — the messy build
Days 1–3: Build an agent or RAG system over genuinely messy data — inconsistent schemas, missing fields, several date formats. Not a clean tutorial dataset.
Days 4–5: Add an eval harness. You will be asked how you knew it worked and this is the answer.
Week 2 — the scoping muscle
Days 1–3: Take three one-line business problems and talk each through for 45 minutes to a recording. Mark the timestamp where you first proposed a solution.
Days 4–5: Write your enterprise-integration whiteboard: where data lives, how auth flows, what happens to PII, the latency budget, who pays for tokens.
Week 3 — the stories
Days 1–2: Three 'they asked for X, the real problem was Y' stories with the discovery moment in each.
Days 3–4: Rehearse the bad-news conversation out loud, with the replan in hand.
Day 5: Prepare the vertical you are targeting — learn its vocabulary and two of its constraints.
The book for this field
LLM Systems Engineering
Seven chapters on the engineering underneath applied AI: the roofline and KV-cache arithmetic that explain every serving decision, quantisation with the evidence, serving stacks and the prompt-cache traps that cost real money, retrieval internals and whether long context replaced RAG, agents at the token level, evaluation with actual statistics, and what won in post-training.
Cross-cutting
Skills every archetype tests
These are shared across every field on this site — the same question asked in different vocabulary — so preparation here compounds rather than being spent once. The book above is specific to your field.
Practice the 22 AI questions → · Back to Artificial Intelligence →