FDE in Action: From a One-Line Brief to a Live AI Page — Full Workflow Breakdown

What Is an FDE? The "Last-Mile" Expert of AI Deployment

As AI capabilities advance at a breakneck pace, a new role is quietly rising to prominence — the FDE (Forward Deployed Engineer). A live Bilibili masterclass demonstrated the full value of this role through a real-time, zero-to-deployment walkthrough.

In one sentence: an FDE is someone who goes to the client's site, armed with AI tools, and turns a vague requirement into a system that actually runs in production. They're not researchers building models — they're engineers making models work inside real client businesses.

Job market data tells a clear story: FDE-related postings on LinkedIn have reached 5,330 openings, with year-over-year hiring growth of 729%. Salaries in China range from ¥10K to ¥100K per month, with major tech companies typically offering ¥30K–¥70K. The signal is unmistakable — AI deployment has a people shortage.

Why Is Demand for FDEs Suddenly Outpacing Supply?

The instructor cut straight to the point: AI deployment has never been bottlenecked by model capability — it's always been the "last mile." Specifically, four problems consistently block AI from landing in production, and none of them can be solved by AI alone:

• Vague requirements: Users want AI to boost efficiency but can't articulate what they actually need
• Dirty data: Real production data has inconsistent formats, missing fields, and conflicting definitions
• Siloed systems: Existing CRM, ERP, and OA platforms are hard to integrate
• No one owns delivery: After launch, nobody wants to be accountable for outcomes

OpenAI's own job description frames the FDE role as: "Own the full arc: Discovery, Technical Scoping, System Design, Building" — end-to-end ownership from requirement discovery through production launch.

What FDE Is NOT: Three Key Distinctions from Pre-Sales, Ops, and Outsourcing

Understanding the FDE role requires knowing what it isn't. The instructor drew three important lines:

FDEs are not pre-sales. Pre-sales reps pitch solutions and hand things off after signing. FDEs start their real work after the contract is signed and are accountable for results. As Tencent's job posting puts it: "Lead deployment delivery and go-live support."

FDEs are not operations engineers. Ops keeps systems from going down; FDEs decide what the system should look like and what it should do — they make "independent technical architecture decisions." Ops reacts to problems; FDEs prevent them at the source.

FDEs are not on-site outsourcing staff. Outsourcing contractors bill by person-day and follow client instructions. FDEs are accountable for business outcome metrics. The instructor was blunt: the career ceiling is completely different — selling your time versus owning results leads to vastly different futures.

Interestingly, FDEs come in two forms: on-site FDEs who embed with clients, and internal FDEs who drive AI adoption within their own companies (some organizations call this role AIBP). That means transitioning to an FDE doesn't require changing jobs — you can become the person inside your current company who turns leadership's ideas into AI-powered reality.

FDE in Practice: From a One-Line Brief to a Live Public URL

The most compelling part of the session was a live demo using Tencent Work Buddy (with the HY3 free model), walking through a complete FDE delivery workflow in real time. The project: build an "FDE Role Radar" page that combined 28 real job requirements scraped from four major tech company career sites with live audience voting data, all displayed on a single page.

The entire process condensed into seven steps, each mapping to a core FDE work pattern.

Step 1: Clarify Requirements — Don't Touch the Code Yet

The instructor stressed that the first FDE move is never to start building — it's to clarify the requirement: who will use this, what problem it solves, what's explicitly out of scope, and three verifiable acceptance criteria.

There was a clever design in the prompt: the AI is not allowed to keep asking questions, but every uncertainty must be written as an explicit "assumption" and logged. The instructor explained why: when given free rein to ask questions in a dry run, the AI once "thought for half an hour without producing a single word" — a pure deadlock. But without any buffer to make assumptions, it would output something that missed the mark entirely. That assumption log becomes exactly the list of items an on-site FDE needs to walk through with the client one by one.

Steps 2–3: Define the Contract, Handle Dirty Data

Step 2 was "locking the contract" — handing the AI the documentation for two voting APIs and plain-language descriptions of six competency options, then having it produce a PRD: a JSON schema for job requirements, field definitions, a mapping table, and a dedicated section for anything it couldn't confidently classify. The instructor called this step "the most representative of day-to-day FDE work."

Step 3 — handling dirty data — was the most grounding moment. Four major tech companies use completely different names for the same role: Tencent calls it "AI Front-Line Deployment Engineer," Alibaba Cloud says "Forward Deployed Engineer," Ant Group uses "Client Technical Solution Architect (FD Direction)," and others simply call it "Delivery Center." The instructor's warning: if you only search for "FDE" when job hunting, you'll miss a massive number of identical roles under different titles — recognizing the same job in different disguises is itself a job-search skill.

Data granularity was also inconsistent — skill requirements were scattered across job responsibilities, qualifications, and bonus criteria sections, requiring human review. Why be this rigorous? Because in real delivery, clients will always ask "where did this data come from?" — that's what auditability means.

Steps 4–7: Build, Self-Test, Deploy, Sign Off

For step 4 — building a single HTML page — the instructor set four hard acceptance criteria, each drawn from FDE experience: API failures must have a visible fallback (never invisible), silent degradation is strictly forbidden, zero votes must still display properly, and the page must show the sample size N without showing percentages (given the limited sample).

The live demo hit a real snag — the first build had a genuine bug. The instructor was unfazed: "Running into problems is the real situation. It actually shows the value more clearly." After one round of fixes, the page successfully displayed voting data from 24 respondents.

The step 5 "self-test and critique" phase exposed the limits of the free model — it churned without producing output for an extended period, and the instructor cut it off decisively. This illustrated a practical reality: AI tools vary significantly in capability, and real FDE work requires fluidly switching between Claude Code, Codex, Trae, Work Buddy, and others depending on the task.

For deployment, since Work Buddy's sandbox couldn't execute remote cloud commands, the instructor manually pushed the output to a remote server with two commands, generating a publicly accessible URL. Audience members opened it on their phones and saw exactly the same page as the local version. A one-line brief had become a live, publicly accessible project.

Who Should Learn FDE? Why Everyone Will Eventually Be One

The instructor offered a clear breakdown of who fits the FDE path:

• Developers, QA engineers, and ops professionals: Already work with code — just need to develop the client-facing perspective
• Pre-sales, solutions architects, and project managers: Already understand requirements — what's missing is hands-on delivery capability (which AI now provides)
• Traditional IT integration professionals: Already do delivery work — that's a natural advantage
• Freelancers and founders: Already know what users want
• Non-tech industry professionals in pharma, logistics, energy, finance, etc.: Domain expertise is the ace up your sleeve

For fresh graduates and career changers with no technical background, the instructor was refreshingly honest: you can learn this and get started, but FDE is a role that genuinely runs on deep industry experience and judgment — it can't be fast-tracked in a few days and takes multiple real-world delivery cycles to develop.

The core thesis was this: everyone should eventually become an FDE. The instructor's plain-language framing: FDE is fundamentally about "one person representing an entire team." What used to require four or five people working together can now be done by one person with AI. That also explains why FDE compensation is high: "You're doing the work of four or five people. Your employer pays you the equivalent of three. Everyone wins."

In an environment where AI is disrupting every role, the path to staying relevant is to operate across multiple functions. You don't have to hold an FDE title — but you need FDE capability: the end-to-end ability to move from requirement conversation to final delivery. That's the core competitive advantage of the future.