Direct answer: FDE = Forward Deployed Engineer (Forward Deployed Engineer); in this article it does not refer to full disk encryption. Core capabilities include software and data, AI applications, production reliability, customer discovery, field delivery, and communication influence.
Learning Objectives
After completing this chapter, you should be able to:
- Identify the common capabilities of FDEs from first-hand job requirements, rather than copying any single company's tech stack;
- Take inventory of your own engineering, AI, production, discovery, delivery, and communication skills across six dimensions;
- Score yourself based only on viewable outputs, distinguishing between "heard of it," "have done it," and "have delivered it under real constraints";
- Identify the three gaps that most affect your target role, and rewrite them as near-term actions with clear acceptance criteria.
Capability Is Not a Tool List
Heading "Capability Is Not a Tool List"
The relevant job titles and business contexts at Palantir, OpenAI, and Anthropic are not identical, but their first-hand job postings repeatedly signal the same things: going deep into customer problems while building with your own hands; moving fast while making systems work reliably in real environments; handling technical details while helping different roles understand trade-offs and adopt solutions.
Therefore, this tutorial does not treat "how many frameworks you know" as the core metric, but instead uses six capability dimensions that can be verified through project evidence.
The Six-Dimension Capability Model
Heading "The Six-Dimension Capability Model"
| Capability dimension | What you actually need to do | Examples of acceptable evidence |
|---|---|---|
| Software and data engineering | Write maintainable code, design APIs, handle data, and complete integrations using Git, testing, and containers | Code repositories, automated tests, API documentation, data models, deployment records |
| AI application engineering | Understand model capabilities and limitations, design retrieval, tool calling, AI agent, or other solutions per scenario, and establish evals | Eval datasets, baseline comparisons, failure cases, prompt and model version records |
| Production and reliability | Handle identity and permissions, security, observability, performance, cost, failures, and rollbacks | Threat models, log/metric dashboards, load tests, incident retrospectives, runbooks |
| Customer discovery and product judgment | Interview stakeholders, map workflows, identify root causes, define success metrics, and control scope | Interview guides, flowcharts, problem statements, metric trees, scope and assumption lists |
| Field delivery and adoption | Iterate in uncertain environments, integrate with customer systems, train users, and complete handover | Pilot plans, release records, user feedback, training materials, handover checklists |
| Communication and influence | Explain the same decision to engineering, business, and management audiences, and drive alignment on risks and trade-offs | ADRs, weekly status reports, demo recordings, decision minutes, cross-role feedback |
AI application engineering is especially important for Forward Deployed AI Engineers, but not all FDEs center on large models. Choose weights based on the target role first; don't force trendy technologies into every project.
Grade with Evidence, Not Gut Feeling
Heading "Grade with Evidence, Not Gut Feeling"
Each dimension uses a 0–3 scale. What is being rated is your currently provable experience, not your potential:
| Level | Judgment criteria | Typical statement |
|---|---|---|
| 0 — No evidence | Understands the concept, but has no viewable output | "I've watched related tutorials" |
| 1 — Guided practice | Completed exercises following explicit steps and can explain the basic approach | "I followed the task to completion and recorded the results" |
| 2 — Independent delivery | Independently handled open-ended problems, made trade-offs, and passed acceptance | "I defined the solution, implemented it, tested it, and explained the trade-offs" |
| 3 — Real constraints | Continuously running, observing, and improving under real user or production constraints | "People use it; I have metrics, incident records, and iteration evidence" |
Having no public repository doesn't mean having no evidence. Sanitized architecture diagrams, metrics, decision records, demos, and retrospectives can also demonstrate capability; for anything that cannot be disclosed, clearly state the confidentiality boundary—never fabricate.
The Minimum Viable Combination for Getting Started
Heading "The Minimum Viable Combination for Getting Started"
For learners with "basic programming ability and zero FDE experience," set the initial targets as:
- Software and data engineering at 2: able to independently complete a small but complete service or application.
- Customer discovery and product judgment at 1–2: able to define problems and acceptance metrics through interviews or scenario research.
- Production and reliability at 1: have at least done deployment, logging, basic permissions, and rollback drills.
- Communication and influence at 1–2: able to explain choices, results, and risks in concise documentation.
- Then, depending on the target role, strengthen AI application engineering or specific industry integration skills.
This is not a hiring bar, nor does reaching these scores mean you'll be hired. It is merely a starting sequence to avoid "learning lots of AI concepts without ever delivering a complete system."
How to Find Your Real Weak Spots
Heading "How to Find Your Real Weak Spots"
Don't compute a seemingly precise overall average score. FDE hiring is more like a constraint system: a 0 in one key dimension may matter more than an extra point in other dimensions.
Judge in the following order:
- Pick 2–3 real target roles and mark which capability dimension each responsibility belongs to.
- Give higher priority to dimensions that are frequently and explicitly required.
- Attach evidence to each of your scores; if you can't provide evidence, downgrade.
- Find the bottleneck that prevents you from completing end-to-end delivery, rather than picking your favorite technology to keep learning.
- Turn the bottleneck into a task that can produce a portfolio piece within two weeks, then re-score.
For example, "not knowing Kubernetes" may not be your first weak spot; if you've never defined success metrics, completing a user interview, a flowchart, and acceptance criteria first may be closer to the core work of the target role.
Verifiable Output: Personal Gap Table
Heading "Verifiable Output: Personal Gap Table"
Create a new
fde-skill-gap.md
, copy and fill in:
# FDE 能力差距 ## 目标岗位| 公司/职位 | 一手岗位 URL | 高频要求 || --- | --- | --- || | | | ## 证据评分| 能力 | 当前等级（0–3） | 证据链接或说明 | 目标等级 | 下一项可验收动作 || --- | --- | --- | --- | --- || 软件与数据工程 | | | | || AI 应用工程 | | | | || 生产与可靠性 | | | | || 客户发现与产品判断 | | | | || 现场交付与采用 | | | | || 沟通与影响力 | | | | | ## 优先补齐的三个缺口1. 缺口、为何优先、截止日期、产物2. 缺口、为何优先、截止日期、产物3. 缺口、为何优先、截止日期、产物
Acceptance criteria:
- Map at least 2 first-hand target roles that are still accessible;
- All six scores have evidence links or clear explanations; "feeling good about yourself" cannot substitute for evidence;
- The three priority gaps must be traceable to target roles, not to generic trending-skill lists;
- Each next-step action can be completed in about two weeks and produce viewable documentation, code, tests, metrics, or demos;
- Invite an engineering peer to review using the same rubric, and record scoring disagreements and reasons for revisions.
Common Pitfalls
- Treating the number of languages and frameworks as capability. What problems you've independently solved is more persuasive than how many dependencies you've installed.
- Only filling AI gaps, not engineering gaps. A successful model call is just the beginning; data, permissions, evals, observability, and failure handling determine whether a system can reach production.
- Equating communication with public speaking. FDE communication also includes uncovering implicit constraints, writing decisions clearly, managing expectations, and driving adoption.
- Waiting until all dimensions reach 3 before applying. Target roles, levels, and team needs differ; honest and clearly documented growth evidence is more credible than a fake perfect score.
- Assuming certificates can replace deliverables. Certificates prove the learning process, but alone they cannot prove you can deliver results in ambiguous situations.
- Dressing up projects with fictional users or metrics. It's fine to clearly state a project is a simulation; don't present test data as real business results.
Next step: take your completed gap table into The FDE Learning Path, and choose your training order based on your target role and evidence gaps.
Authoritative Sources
- Palantir — Forward Deployed Software Engineer
- Palantir — Forward Deployed AI Engineer
- OpenAI — Forward Deployed Engineer
- Anthropic — Forward Deployed Engineer
Verification date: 2026-08-19. The six-dimension model and the 0–3 scoring method are original teaching designs of this tutorial based on the first-hand job signals above, and are not the official competency frameworks or hiring commitments of the recruiting companies.