Sources and Factual Boundaries

Key Takeaways First

Title "Key Takeaways First"

This tutorial fixes FDE as Forward Deployed Engineer, with the Chinese term covering both "前沿部署工程师" and "前线部署工程师". The course prioritizes official company recruiting pages, official engineering blogs, and technical specifications; media reports, course promotions, and community posts can only serve as leads and cannot alone support job facts.

Verification baseline for job and foundational sources: 2026-08-19. The Capstone evals, security, and operations source links added in weeks 21–24 were re-verified on 2026-09-12; the Google SRE page retains its existing verification from 2026-08-19, as this online recheck timed out, so it is not written as re-confirmed. External content and versions may still change.

To return from sources to the tutorial, read in order Job and Role Definitions, Capability Matrix, Learning Path, and Hands-on Projects.

Job and Role Definitions

Title "Job and Role Definitions"

AI Applications and Production Delivery Materials

Title "AI Applications and Production Delivery Materials"

• OpenAI Docs — Agents SDK: Agents, tools, state, approvals, and observability entry points.
• OpenAI Docs — Evaluation best practices: Evals goals, datasets, metrics, and principles for continuous evals.
• OpenAI Docs — Production best practices: Safety, scaling, cost, latency, and production environment principles.
• Anthropic — Building Effective Agents: The boundary between workflows and agents, and common composition patterns.
• Model Context Protocol: Protocol concepts and implementation entry points for MCP.
• OWASP GenAI Security Project: Key security risks for generative AI applications.

Customer Discovery and Business Judgment Materials

Title "Customer Discovery and Business Judgment Materials"

• GOV.UK — Start by learning user needs: Start from users, current behavior, and problems; do not treat solutions proposed by stakeholders as user needs.
• GOV.UK — How the discovery phase works: Questions, scope, constraints, data, and stop decisions in Discovery. Its public service context cannot substitute for enterprise procurement and business judgment.
• Google PAIR — Identify user needs and AI strengths: Identify user contexts, map existing workflows, and compare AI, rule-based, and human approaches.
• Palantir — A Day in the Life of an FDSE: Team practices in customer collaboration, domain learning, engineering, and product feedback; also serves recruiting and promotional purposes.
• Ramp — Forward Deployed Engineering: Team practices in continuous scope judgment, direct user contact, and measuring work by customer outcomes; also serves company culture and recruiting promotional purposes.

The "two-layer dialogue", "five business questions", and evidence translation board in this tutorial are original teaching tools synthesized from the above materials, not a unified FDE methodology published by any organization.

Capstone Evals, Security, and Operations Materials

Title "Capstone Evals, Security, and Operations Materials"

• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks: The original research paper on RAG; the course's citations, refusal to answer, permissions, and thresholds are subsequent instructional design, not requirements of the paper.
• NIST SP 800-162 — Guide to Attribute Based Access Control: A reference for making access control decisions based on subject, object, action, and environment attributes.
• Model Context Protocol Specification: MCP Host, Client, Server, capability negotiation, and protocol contracts; when verified on 2026-09-12, latest resolved to the 2026-07-28 version.
• MCP Security Best Practices: A reference for authorization, tokens, sessions, proxies, and remote server risks.
• OWASP Authorization Cheat Sheet: A reference for deny by default, checking permissions on every request, and authorization testing principles.
• NIST AI RMF Generative AI Profile: A reference for generative AI risk identification, measurement, and governance.
• OWASP Top 10 for LLM Applications: A reference for prompt injection, sensitive information, supply chain, data or model poisoning, permissions, and excessive agency risks.
• OpenTelemetry Generative AI Semantic Conventions: A reference for model and agent telemetry fields; when implementing, you must check the stability level of specific fields.
• RAGAS: Automated Evaluation of Retrieval Augmented Generation: a research reference for RAG eval dimensions; the course's data distribution, thresholds, and release decisions are not industry standards given by this paper.
• Google SRE: Monitoring Distributed Systems: a reference for operational judgment on latency, traffic, errors, saturation, and user-facing symptoms; the 2026-08-19 verification baseline is retained, and this online recheck timed out.

These sources support the protocols, threats, observability, and eval methods; they do not support the fictional Beichen case, sample sizes, thresholds, or pilot results in this course. All course numbers must be labeled as instructional design; real projects should be re-approved by authorized roles based on their own tasks and risks.

Search and Generative Citation Principles

Heading "Search and Generative Citation Principles"

To serve both traditional search and generative search, this tutorial follows these writing practices:

• Each page first presents conclusions that can be cited independently, then expands on evidence and steps.
• Abbreviations are given with their full names, Chinese aliases, and clear context on first appearance.
• Facts, recommendations, and author judgments are expressed separately.
• Sources and verification dates are noted for roles, salaries, quantities, and product versions.
• Pages use stable URLs, static HTML, canonical tags, sitemaps, and structured data.
• llms.txt and llms-full.txt are provided, but these non-mandatory standards are not described as guarantees of indexing.

OpenAI's official documentation states that the crawler used for ChatGPT search display is OAI-SearchBot, which is independent of GPTBot, used for model training control. The site root's

robots.txt

should allow OAI-SearchBot; whether to allow GPTBot should be decided separately by the site owner based on content policy. See

。

Practices We Will Not Adopt

Heading "Practices We Will Not Adopt"

• No republishing paid courses, leaked materials, or wholesale copying of third-party articles.
• No using unverifiable claims like "demand growing dozens of times" or "million-dollar salaries" to create urgency.
• No fabricating certificate endorsements, student outcomes, partnerships, or project metrics.
• No treating search autocomplete suggestions as search volume or market size.
• No creating duplicate-content pages for "Forward Deployed Engineer" and "Forward Deployed Engineer".

If a source becomes invalid or facts change, you can submit the link and verification date via a GitHub Issue.