Skip to content

About FDEs

FDE Glossary

Common terms in FDE and AI adoption work, explained as plainly as possible.

FDE (Forward Deployed Engineer)
An engineer deployed at the customer's site to turn AI and software capabilities into systems the customer's business actually uses and that show results. The work runs from requirements interviews through launch and delivery.
AI adoption
Applying AI capabilities to specific business processes with measurable results. A demo that runs does not count. Adoption means the business keeps using it.
Use-case selection
Among the many ideas a customer proposes, judging which use case has high value, the necessary data and conditions, and controllable risk, so it is worth doing first. It is one of an FDE's core skills.
Scope creep
Requirements keep growing during a project, so the timeline and cost spiral out of control. The remedy is an explicit "out of scope" list in the proposal and managing every change.
PoC (Proof of Concept)
A small, short pilot that checks whether a solution is feasible. A good PoC has clear success metrics and, at the end, answers whether to invest further.
Baseline
Current-state data from before the project starts, such as today's average response time or error rate. Without a baseline, you cannot prove the project brought an improvement.
Large Language Model (LLM)
A language model trained on large amounts of text that can understand and generate natural language. FDEs usually do not train LLMs. They combine them with the enterprise's knowledge and systems.
Prompt
Instructions and context given to a model. A system prompt usually contains the role, the task, the boundaries, the output format, and what to do when the model is unsure.
RAG (Retrieval-Augmented Generation)
Retrieve relevant content from the enterprise knowledge base first, then have the model answer based on that content. It lets the model use the enterprise's own up-to-date knowledge and reduces fabrication.
Knowledge base governance
Deciding which documents may enter the knowledge base and keeping them accurate and current. This includes removing outdated and conflicting content, stripping sensitive information, and assigning an owner.
Agent
An AI system that decides its own steps toward a goal, calls tools and completes the task. Compared with single-turn Q&A, an agent can query systems and take actions, so it needs stricter permissions and fallback design.
Tool use / Function calling
Having a model request calls to external functions in an agreed format, such as querying an order or sending an email. It is the main way models connect to business systems.
Hallucination
The model generates content that looks plausible but is incorrect or unsupported. Common controls are limiting knowledge sources, requiring the model to state uncertainty, and checking with an evaluation set.
Evaluation set
A set of test inputs and pass criteria for checking an AI system. A good evaluation set covers normal questions, edge cases, questions outside the knowledge base, and adversarial inputs.
Held-out test set
Test cases that the people building the system cannot see in advance. It checks how the system performs on questions it has not seen and prevents tuning only to known questions.
Prompt injection
Instructions hidden in user input or external content that try to make the model ignore its original rules, for example to leak internal information. Systems facing external users must be tested against this kind of input.
Fallback
How the system handles cases where the AI cannot or should not answer, such as saying so clearly, handing off to a human, or collecting a contact. The fallback design determines how bad the consequences are when the system fails.
Human-in-the-loop
Keeping human review or confirmation at key steps: the AI drafts and recommends, and people decide. Suited to scenarios where errors are costly, such as external price quotes or medical advice.
On-premises deployment
Deploying models and applications on the customer's own servers or intranet so the data never leaves the customer's environment. Common in manufacturing, finance, government and healthcare.
Data egress
Data leaving an environment the enterprise controls, for example when sent to an external model service. When personal information or trade secrets are involved, the scoping stage must confirm whether this is allowed and what approvals are needed.
Gradual rollout
Launching first to a small scope (some stores, some hours, some users), observing results and problems, then expanding step by step. This keeps the impact of errors contained.
Rollback
Reverting to the previous state when a serious problem appears after launch. The launch plan should state in advance when to roll back, who decides, and how to do it.
ROI (Return on Investment)
The ratio of a project's return to its investment. FDEs measure it with the customer's own data, such as hours saved, churn reduced, or downtime shortened.
Learning is the exam
ChinFDE's learning model: learners complete interviews, proposals, builds and retrospectives in a simulated customer project. The outputs they leave along the way are the basis for assessment; there is no separate written exam.