The role AI companies are scrambling to hire for, with no university major—someone has open-sourced the entire learning path

There's a role that AI companies are fighting to hire for, but universities don't have a matching major, and training programs can't offer any decent courses for it. It's called FDE, Forward Deployed Engineer.

Amit Shekhar recently open-sourced a complete roadmap to becoming an FDE on X, hosted on GitHub, with one blog post per topic, all free.

Let's first clarify what this role does. AI Engineers are responsible for building the product. FDEs are responsible for fitting the product into a specific customer's production system, getting their data, workflows, and permissions to work, and then bringing the lessons learned back to the product team. The former leans toward building; the latter leans toward applying.

It sounds like pre-sales, but in reality you have to write production code, build RAG pipelines, integrate APIs, and sit in customer meeting rooms translating a sentence like "We want to use AI to optimize customer service" into something that can be accepted and signed off. So it demands breadth: LLMs, RAG, agents, evaluation, inference optimization, security, system design—you need to know enough about each to make on-the-spot judgments.

The problem is that this skill set doesn't have a corresponding major. Computer science departments can't teach how to make technical decisions in front of a customer, and bootcamps can't teach how to use evals to prove a system is worth the money. Amit's roadmap fills this gap.

14 steps, arranged according to the actual lifecycle of an FDE project: discover requirements, learn the toolbox, build a prototype, prove value, harden, deploy. Under each step hang several blog posts, each explaining one concept, most with videos. Let's highlight a few.

Step 2 covers LLM internals, including why hallucinations happen, why long documents get forgotten, and why responses are slow. These are the most frequently asked questions on customer sites. If you can't answer them, trust is gone.

Step 5 is entirely about RAG. From vector databases and ANN search to hybrid retrieval, rerankers, chunking strategies, and then GraphRAG and RAG without vectors. RAG is what FDEs deliver most often, so this section is the thickest.

Step 8 is evaluation and observability. This step is easy to skip, but it determines whether you can get the acceptance signature. No matter how good the system is, if you can't show numbers, the customer won't sign off.

Step 13 is the finale: designing a real-time voice AI agent. Voice scenarios compress models, agents, tool calling, latency, streaming, and system design all together. Endpoint detection, barge-in, first-audio latency, TTS chunking—it breaks each one down. If you can explain this pipeline clearly, you'll feel confident when you encounter system design questions in interviews.

The last step of the roadmap is interview preparation. Amit says FDE interviews test three things: writing code with AI coding agents, AI system design, and customer scenarios. The third step is the trickiest: you're given a vague business problem, and they watch what questions you ask, how you define success, what solution you choose, and whether you can explain it clearly.

Every link in the repo points to an independent blog post, free to read, no registration required. The author is the founder of Outcome School and says it will be continuously updated.

In the comments, a reader asked: I just learned about FDE today—should I sign up for a paid course? First go through this roadmap, then decide whether to spend money.

Repo link: forward-deployed-engineer-roadmap