Engineeringhot

Forward Deployed Engineer

Also hired as: FDE · Forward Deployed Software Engineer · Applied AI Engineer (Customer)

Ship frontier AI on real enterprise data — the engineer the customer actually meets.

US salary · 2026
$160k – $350k+ total comp
Palantir-style FDE roles and consultancy equivalents sit in the lower half; frontier labs pay MTS-adjacent packages at the top, equity-heavy. Location and travel expectations move the number more than the title does.
Typical background
Full-stack or backend engineers who like customers; ex-consultants who learned to ship; solutions engineers who got tired of handing off the fun part.
This roadmap
5 stages · 25 nodes
5 proof-of-work checkpoints
What is a Forward Deployed Engineer?

A Forward Deployed Engineer ships working AI software inside a customer's four walls. Palantir invented the title two decades ago; the AI wave made it one of the hottest engineering jobs in the industry, with OpenAI, Anthropic, Scale, and every serious AI application company building FDE teams. You embed with a customer, learn their actual workflow, and build the agent, RAG system, or automation that runs on their real data — behind their firewall, past their security review, into their P&L. You are the last mile between a frontier model and a business outcome, and the last mile is where the money is.

The role exists because enterprises bought AI faster than they built AI teams. Models are general; businesses are specific. The distance between a slick demo and a system that survives contact with a claims department's fifteen-year-old Oracle schema is measured in engineering, not prompting — and that distance is where most enterprise AI deals die. FDEs close it. That's why the labs went from a handful of forward deployed engineers in 2024 to entire field-engineering orgs by 2026, and why the role commands product-engineer pay with sales-adjacent leverage and visibility.

Don't confuse it with neighboring roles. A solutions architect advises and hands off; you write production code and own the outcome. A consultant bills hours against a deck; you leave running software behind. A product engineer serves a roadmap; your roadmap is whatever unblocks this customer this month — and the patterns you extract feed back into the product as features and eval sets. It suits engineers who find pure product work too slow and pure consulting too hollow.

What you'll actually do
  • Embed with a customer's team for two to six weeks: map one workflow end to end, then ship a working system against their real data before the engagement ends.
  • Run discovery sessions with domain experts — claims adjusters, paralegals, underwriters — and turn tribal knowledge into specs, eval sets, and acceptance criteria.
  • Build the unglamorous integrations that close deals: Salesforce, ServiceNow, SharePoint, Snowflake, and the legacy SQL nobody documented.
  • Harden a pilot into production: SSO, VPC deployment, audit logs, rate limits, cost ceilings, and an on-call story.
  • Design evals with the customer's own experts so 'it works' is a number both sides signed off on.
  • Demo weekly to stakeholders from analyst to CIO, and defend the system in security review.
  • Scope engagements with account executives — estimate feasibility honestly before the contract is signed, not after.
  • Carry field patterns back to product and research as feature requests, eval data, and 'stop promising this' memos.
This role fits you if
  • Shipping something imperfect today beats polishing something perfect for next month, and that ordering feels obviously right to you.
  • You can interview a claims adjuster at 10am and debug a Kubernetes ingress at 4pm without visibly changing gears.
  • Ambiguity reads as room to move, not missing requirements.
  • You pick up a new stack, schema, or industry vocabulary in days, because the customer doesn't care what you knew last week.
  • Regular customer time — on site or on calls — energizes you rather than draining you.
  • You want your work visible in a customer's revenue, not just in a sprint report.
The toolbox
PythonTypeScriptClaude CodeRAG pipelinesMCPEvalsPostgresDocker & KubernetesTerraformSnowflake / DatabricksSSO / SAML / OktaPrompt engineeringVector searchEnterprise APIs (Salesforce, ServiceNow)
The roadmap — 5 stages, 25 nodes
Stage 1

Full-stack + LLM foundations

Weeks 1–6

FDEs are generalists under pressure. Build the floor: ship a full-stack app fast, and understand what a model actually is before you sell one.

Get dangerous across the stackCore
Python and TypeScript, a REST API, Postgres, and one deploy path (Docker on a cloud box or Vercel). You don't need depth everywhere — you need to never be blocked by a layer you can't touch.
How LLMs actually workCore
Tokens, context windows, sampling, and why models fail the way they do. Every customer conversation you'll ever have routes through this mental model.
Prompting as engineeringCore
Output contracts, few-shot examples, delimited inputs. On site you'll write prompts in front of skeptical experts — they need to work on the second try, not the twentieth.
Adopt an agentic coding workflowElective
Learn Claude Code or an equivalent agent well enough that a one-week build fits in two days. FDE timelines assume this leverage now.
Checkpoint: ship a full-stack app with one LLM feature✓ Checkpoint
Auth, database, one genuinely useful LLM feature, deployed to a public URL — built end to end in under two weeks. This is the minimum unit of FDE work.
Stage 2

Applied AI patterns

Weeks 7–14

The four systems you will build at every customer, in some combination: retrieval, agents, integrations, and the evals that prove they work.

RAG, properlyCore
Chunking, embeddings, hybrid search, reranking, citations. Then the part demos skip: what happens when the documents are scanned PDFs with tables.
Agentic systemsCore
Tool use, loops, and the honest judgment call: when an agent beats a pipeline and when it's an expensive way to be nondeterministic.
MCP and systems of recordCore
Model Context Protocol is how you'll wire models into customer tools without bespoke glue every time. Build one MCP server against a real API.
Evals as a deliverableCore
Golden sets, LLM-as-judge, and why 'it looked good in testing' isn't a sentence you can say to a CIO. At customer sites, the eval is often the artifact that closes the deal.
Checkpoint: RAG over 500 real documents✓ Checkpoint
Ship a working RAG service over 500+ real documents — public filings, a documentation corpus, a wiki export. A stranger can query it, answers carry citations, and you can quote its retrieval hit rate from a written eval.
Stage 3

Enterprise reality

Weeks 15–24

The gap between a demo and a deployment is security review, permissions, and cost math. This stage is why FDEs get paid.

Enterprise data plumbingCore
Stand up a free Salesforce dev org or SharePoint sandbox and pull real objects through their APIs. Learn what enterprise data actually looks like: permissioned, denormalized, and half-empty.
Security and deployment constraintsCore
SSO/SAML, VPC deployment, data residency, prompt injection, PII handling. Practice answering a vendor security questionnaire — you will face dozens.
Cost and latency engineeringCore
Token budgets, caching, model routing, batch vs realtime. Enterprises sign annual contracts — you need to project a year of inference cost without flinching.
Go deep on one enterprise ecosystemElective
Pick Salesforce, ServiceNow, or Epic and learn its object model and integration idioms. One deep ecosystem makes you instantly credible in half your future engagements.
Checkpoint: make your RAG system enterprise-grade✓ Checkpoint
Add SSO login, per-user document permissions (two users see different answers to the same query), and an audit log to your Stage 2 system. Boring on purpose — this is the work that closes deals.
Stage 4

Field craft

Months 7–9

The customer-facing half of the job: extract the real workflow, demo like it matters, and scope work you can actually deliver.

Discovery interviewsCore
Sit with a domain expert and extract a workflow: inputs, outputs, exceptions, and the spreadsheet nobody mentioned. Write it up as a spec with acceptance criteria the expert signs.
Demo craftCore
Build a demo on the customer's own data within days and narrate business value, not architecture. Rehearse failure: what you say when the model gets it wrong live matters more than when it's right.
Scoping and expectation managementCore
Write a pilot-to-production plan with success metrics agreed before the build starts. Underpromising is a survival skill: every oversold capability becomes your personal debt.
Learn one vertical's languageElective
Insurance claims, legal review, pharma, logistics — pick one and learn its acronyms, systems, and regulatory constraints. FDE staffing is increasingly vertical.
Checkpoint: run a real engagement✓ Checkpoint
Find a real 'customer' — a local business, a nonprofit, another team at work. Run a discovery call, build for one week on their actual data, deliver a live demo, and write up before/after metrics they agree with.
Stage 5

Break in and get hired

Months 9–12

FDE interviews test exactly what the job is: build under ambiguity, in front of people. Package your proof and practice the loop.

Portfolio as outcome storiesCore
Write your three best projects as case studies: the problem, the constraint, what shipped, the measured result. FDE hiring managers read for outcomes and speed, not tech lists.
The FDE interview loopCore
Expect live coding, system design with deliberately vague requirements, and a customer role-play where the interviewer changes their mind mid-session. Practice thinking out loud while building.
Target the right employersCore
Frontier labs, AI-native application companies (Sierra, Harvey, Glean, Distyl), Palantir, and consultancy AI arms each weigh the engineering/customer mix differently. Tailor your story per type.
Publish field-pattern teardownsElective
Write about deployed-AI patterns — why pilots die, how permissioned RAG works. Public thinking is how FDE recruiters find you between roles.
Checkpoint: pass a full mock FDE loop✓ Checkpoint
Run a complete mock loop with a peer: timed build from a vague prompt, then a role-played customer demo where requirements shift mid-demo. Record it, review it, and fix the two weakest moments before real interviews.
Build your portfolio

Nobody hires a Forward Deployed Engineer off a certificate. They hire off proof. Ship these and put them where people can click them:

01
The two-week full-stack ship
A deployed app with auth, Postgres, and one genuinely useful LLM feature — built end to end in under two weeks, live at a public URL a stranger can sign into.
Proves: You can deliver the minimum unit of FDE work on an FDE timeline.
02
Messy-data-to-insights pipeline
Take a hostile real-world export — broken headers, mixed encodings, half-empty columns — and ship a pipeline that cleans it and answers questions over it with citations. Commit the ugly input file to the repo.
Proves: You stay useful when the data looks like enterprise data, not like a demo.
03
Permissioned RAG over 500 real documents
Retrieval with citations over 500+ real documents, plus SSO login, per-user permissions (two users get different answers to the same query), and an audit log. Quote its retrieval hit rate from a written eval.
Proves: You know the distance between a demo and a deployment — and you've closed it once.
04
An MCP server against a real system of record
A working MCP server over a Salesforce dev org or ServiceNow sandbox that lets a model read and write real objects, with scoped permissions and a README that gets a stranger to a working call in ten minutes.
Proves: You can wire a model into the systems enterprises actually run on.
05
A real engagement, written up
A discovery doc with signed acceptance criteria, a one-week build on a real organization's data, a live demo, and a before/after metrics writeup the 'customer' agreed with. A nonprofit or another team at work counts fully.
Proves: You can run the entire FDE loop — discovery to measured outcome — without supervision.
Position your profile
Headline formula

I embed with customers and ship AI systems on their data — [N] systems from discovery to production, most recently [outcome] for a [industry] team.

Resume bullets to earn
  • Shipped a permissioned RAG system over [N] documents with SSO and audit logging; [X]% retrieval hit rate on an eval the domain experts signed.
  • Ran a discovery-to-demo engagement in [N] weeks: mapped a [claims/contracts] workflow, shipped a pilot on real data, cut cycle time [X]%.
  • Built MCP integrations into [Salesforce/ServiceNow] so agents act on live records behind per-user permissions.
  • Designed acceptance evals with customer SMEs so 'done' was a number both sides signed — then hit it.
  • Projected a 12-month inference budget within [X]% of actuals using model routing, caching, and batch workloads.
Where to be visible
  • Pin 3–4 repos on GitHub with READMEs that lead with the business outcome and a ten-minute run-it-yourself path — hiring managers click exactly one thing.
  • Write field-pattern teardowns: why pilots die, how permissioned RAG works, what security review actually asks. This is how FDE recruiters find you between roles.
  • Publish one engagement case study — problem, constraint, what shipped, measured result — even if the customer was a nonprofit.
  • Give a meetup or lightning talk on 'demo to deployment'; FDE hiring favors people who can visibly hold a room.
  • Contribute a connector or fix to the MCP ecosystem — integration credibility, in public.
What interviews actually test
  • Live build from a deliberately vague prompt, timed. Practice narrating decisions while coding — composure is the thing being graded.
  • System design on messy, permissioned data: legacy schema, partial access, an SSO requirement. Bring your permissioned-RAG design; it maps directly.
  • Customer role-play where the interviewer changes requirements mid-session. Rehearse scoping questions and the art of saying 'not in this pilot' warmly.
  • Behavioral on ambiguity and ownership — prepare stories where you shipped without a spec and measured the result.
  • The scoping exercise: 'the customer wants X in four weeks — what do you commit to?' Underpromise on record, with a written pilot plan.
Who's hiring
PalantirOpenAIAnthropicScale AISierraGleanHarveyDistyl AIDatabricksAccenture & the big consultancies' AI arms
Free courses that map to this role
Related roles
FAQ

Do I need a CS degree to become a Forward Deployed Engineer?

No. FDE hiring runs on proof you can ship under ambiguity and hold a room — a portfolio of deployed systems plus strong communication beats credentials at most companies. Some frontier labs still lean on pedigree for junior hires, but consultancies, AI-native startups, and Palantir-style orgs hire non-traditional backgrounds constantly.

How is an FDE different from a solutions architect?

A solutions architect designs, advises, and hands off to someone else's engineers. An FDE writes and ships the production code personally and owns the outcome. FDE is graded as an engineering role — expect coding interviews — while SA roles usually aren't. Comp bands are typically higher on the FDE side at labs for the same reason.

How much travel does the job actually involve?

It ranges from Palantir-style weeks on the customer's site to remote-first labs where 'forward deployed' mostly means video calls with quarterly onsites. It's set by the employer and the account, not the title. Ask directly in interviews: 'how many days per month on customer site, on average?' and get the answer in writing.

Is FDE a career dead end compared to product engineering?

The opposite, lately. FDEs accumulate the rarest knowledge in the company — what real customers do with the product — which converts into product leadership, GTM leadership, or founding teams. The Palantir FDE alumni network turning into startup founders is the canonical proof.

What do FDE interviews look like?

Usually: practical live coding, a system design round with intentionally underspecified requirements, and a customer-facing role-play or presentation. Some companies add a take-home 'build a demo from this vague ask' exercise. The consistent theme is composure while building under ambiguity — rehearse exactly that.

Start building, today

Every stage above maps to free lessons on this site. No signup, no paywall — open the first course and ship your first checkpoint this week.

Browse the courses →