Forward Deployed Engineer
I'm a forward deployed AI engineer based in Melbourne. I embed with your team, run the discovery myself, and build the MCP servers, agent skills and plugins that actually clear a reviewer's sign-off. Production, with tests, in your environment.
A forward deployed engineer (FDE) is a software engineer who embeds inside a customer's environment and owns an AI system end to end: scoping it, writing the production code that runs in the customer's systems, and keeping it running afterwards.
The role was invented at Palantir in the early 2010s, where it was known internally as Deltas. For roughly the company's first decade there were more forward deployed engineers than product engineers. It existed because intelligence customers could not openly say what they needed, so Palantir sent engineers to sit inside the customer and find out by building.
The 2026 version is the same shape pointed at LLMs, and it is being hired by OpenAI, Anthropic, Google, Databricks, Salesforce and Deloitte. The reason is that the bottleneck moved. It is no longer "can the model do this". It is "can anyone get this into a regulated workflow, with the messy data, the legacy system, the approval gate, and the person who signs their name to the output". That last mile is the job.
Forward deployed software engineer and forward deployed AI engineer (also written FDE AI) are the same role under different labels. Palantir used the software framing first; the AI labs shifted the emphasis to agent and LLM work.
Three lines separate a forward deployed engineer from everything that looks like it.
| Role | Owns the design | Writes production code | Owns the outcome after go-live | In practice |
|---|---|---|---|---|
| Consultant | Yes | No | No | Recommends, hands over a deck, leaves. |
| Solutions Architect | Yes | Proofs of concept only | No | Designs the system, hands the build to someone else. |
| AI / ML Engineer | Product side | Yes | The product, not the customer | Builds models, evals and training infrastructure. Rarely customer-facing. |
| Forward Deployed Engineer | Yes | Yes, in the customer's system | Yes | Scopes it, writes it, ships it, and is measured on what happens after. |
Reported time split is roughly 60% customer-facing, 30% deployment-specific code, 10% internal. A consultant recommends. An architect designs. An FDE is the one whose name is on the thing when it runs on Monday.
01
Discovery one person at a time, in your systems, on live files. Not a workshop, and not a survey. Each session follows a written interview method so the results are comparable and the programme survives being run by someone other than me.
Output is a process definition document, a variant catalogue and an event log. Not a summary email.
02
A process is not automatable until you know where the judgement sits. That means breaking a task down past the step level, to the individual points where a person decides something, and separating the branches a machine can take from the ones it must not.
Every task gets two numbers on the spot: how many a week, and how long each. No number, no business case.
03
MCP servers, sub-agents, agent skills and Claude Code plugins running against real data, in your repository, under your controls. Python and TypeScript, with tests, a changelog and a version number, because an artifact nobody can maintain is a liability.
Source of truth lives in one place, and that rule goes into the project context file so every future session inherits it.
04
Nothing counts as shipped because it demoed well. Before the build starts we agree the bar in writing: a fixed set of representative cases, a required accuracy threshold, a user acceptance threshold, a written specification and a named sign-off. Your bar, set against your risk appetite.
Evaluation frameworks are the part most AI delivery skips. They are why the second and third deployment go faster than the first.
05
In regulated work the boundary is the product. Anything that would write to a regulator, a payment rail or a system of record is preview only by default, and a qualified human reviews and signs before it goes anywhere.
Human in the loop by design, not by policy retrofit after an incident.
06
Engagements are structured so I am still iterating against real cases months after go-live, because the failure modes that matter only appear once the system meets data nobody anticipated.
An FDE is measured on what happens after, which means you do not get to leave at go-live.
Six artifacts, every engagement. All of them live in your repository, under your controls, and none of them require me to stay.
The written record of how a process actually runs, not how the policy says it runs. Includes the variant catalogue, because the exceptions are where automation projects die, and an event log you can measure against later.
Model Context Protocol servers that expose your systems to an agent under your own auth and your own permissions model. Scoped tools, typed inputs, audit logging, and a deliberate decision about which operations are read-only.
Packaged, versioned capabilities rather than prompts pasted into a chat window. Each one has a defined input contract, a test set, and a clear boundary on what it is allowed to touch.
A repeatable test set with a pass threshold, run on every change. This is the artifact that turns "it seemed to work" into a number you can put in front of a risk committee, and the reason later deployments go faster.
The human-in-the-loop boundary, written down before the build: what the system may do unattended, what is preview only, who signs, and what happens when it is unsure. In regulated work this is the deliverable, not the paperwork around it.
Source in your repository, tests, a changelog, a project context file, and a runbook. The measure of a good embedded engagement is that your team can extend the thing after I stop invoicing.
Client work stays confidential. Nothing from a current or past engagement, including internal standards, thresholds, tooling or process detail, appears on this page or in a pitch. That rule protects the next client as much as the last one.
Eight things this work demands, and where mine comes from.
Palantir invented it and called them Deltas. OpenAI runs scoping, validation and delivery phases onsite. Anthropic hires into Applied AI, where the stated deliverables are MCP servers, sub-agents and agent skills. Google covers Cloud and GenAI delivery. Salesforce has advertised an Agentforce FDE for Sydney and Melbourne. Databricks, Deloitte across four Australian cities, and AI scaleups including Lorikeet, ElevenLabs, Ramp and Commure are all hiring the same shape.
| Market | Figure | Note |
|---|---|---|
| United States, average total comp | ~US$238k | Band US$205k to US$486k. Staff level clears US$630k. |
| Frontier labs, mid to senior | US$350k to US$550k | Anthropic’s posted FDE base band is roughly US$280k to US$320k. |
| Australia, senior total package | A$180k to A$250k+ | Smaller market. A handful of titled roles, many more under other labels. |
Treat the title as a weak salary signal. What moves the number is whether you can show a production system you shipped inside someone else's environment.
No certification exists. Anyone selling one is selling a brand, not a credential.
Hiring managers assess evidence: a production system you shipped inside someone else's environment, an evaluation framework you wrote, an MCP server or agent skill you can walk through line by line, and the ability to decompose an ambiguous customer problem out loud. That last one is the signature interview round. It runs 45 to 60 minutes, carries roughly 30% of the weight, and has the lowest pass rate in the loop at around 40%.
The tooling half is learnable, and it is the half I teach. Two of my courses map directly onto the FDE toolchain.
Project context files, skills, sub-agents, MCP integration, and driving real systems from the terminal. This is the toolchain the artifact list in the job postings describes.
View the courseOrchestration and integration patterns, which is the other half of getting an agent to touch real business systems. Taught as an official n8n Ambassador.
View the courseReferral links, no extra cost to you. Full catalogue on the courses page.
A forward deployed engineer (FDE) is a software engineer who embeds inside a customer’s environment and owns an AI system end to end: scoping it, writing the production code that runs in the customer’s systems, and keeping it running afterwards. The role was invented at Palantir in the early 2010s, where it was known internally as Deltas, and has since been adopted by OpenAI, Anthropic, Google, Databricks, Salesforce and Deloitte.
A solutions architect owns the design and typically builds proofs of concept, then hands the build to a delivery team and does not own the outcome after go-live. A forward deployed engineer owns the design, writes the production code inside the customer’s system, and is measured on what happens after it ships. A consultant recommends, an architect designs, an FDE is the one whose name is on the thing when it runs on Monday.
A forward deployed AI engineer is the same role pointed specifically at LLM and agent systems. The deliverables are usually MCP servers, sub-agents, agent skills and evaluation frameworks, built and deployed inside the customer’s environment rather than shipped as a product. The term FDE AI is used interchangeably.
Forward deployed software engineer is the same role under a different label. Palantir used the software engineering framing first, and the AI labs shifted the emphasis to LLM and agent work from around 2024. The defining traits are unchanged: embedded in the customer environment, writes production code, owns the outcome after go-live.
In the United States, average total compensation is around US$238,000, with a band of roughly US$205,000 to US$486,000 and staff level clearing US$630,000. Frontier labs sit around US$350,000 to US$550,000 for mid to senior hires. In Australia the market is much smaller, with senior total packages around A$180,000 to A$250,000 and above. Treat the title itself as a weak salary signal: what moves the number is whether you can show a production system you shipped inside someone else’s environment.
No. There is no recognised forward deployed engineer certification, and anyone selling one is selling a brand rather than a credential. Hiring managers assess evidence instead: a production system you shipped inside someone else’s environment, an evaluation framework you wrote, an MCP server or agent skill you can walk through line by line, and the ability to decompose an ambiguous customer problem out loud.
Build and ship one real system inside a real environment, with tests, a changelog and a written evaluation bar, then be able to explain every decision in it. The tooling half is learnable: project context files, agent skills, sub-agents, MCP integration and orchestration. The harder half is customer discovery under ambiguity, which is the signature interview round, runs 45 to 60 minutes and has the lowest pass rate in the loop.
Yes, but not many under that exact title. Palantir, OpenAI, Databricks, Salesforce, Deloitte and the Sydney startup Lorikeet have all advertised the role in Sydney or Melbourne. A far larger pool of Australian organisations is doing the same work under other labels: AI Solutions Engineer, Agentic AI Lead, AI Delivery Lead, Applied AI Engineer and Head of Intelligent Automation.
Yes. Embedded engagements run across Australia and remote, and typically start with a two to four week process discovery inside your business. The output is a process definition document, a variant catalogue and an event log, followed by the build. Email [email protected] or send a message on LinkedIn at linkedin.com/in/amjidali.
Longer-form analysis on the same work: MCP in the enterprise, production-grade AI development, and why most pilots never ship.
How to build Model Context Protocol servers your security team will actually approve: SSO, audit trails, tenant isolation. The handbook, from an operator shipping MCP.
Read analysisHow engineering teams moved past Copilot novelty to AI-assisted development that ships production code: the patterns, governance, and measurable velocity wins in 2026.
Read analysisMIT says 95% of GenAI pilots never reach production. RAND puts the project-value failure rate at 80%. The survivors all do the same five unsexy things in the same order. Here they are.
Read analysisEmbedded engagements across Australia and remote, starting with a two to four week process discovery inside your business. 30-minute call, no slides.