Why
Two years of AI coding assistants have moved the bottleneck. Producing a working function is no longer the expensive step; deciding that a problem needs retrieval rather than fine-tuning, an agent loop rather than a fixed workflow, or no model at all, is. The video makes that explicit at the top (00:43): when code is cheap, engineering judgment is the differentiator. A roadmap that names the tiers matters for the same reason a syllabus does. It tells a working engineer which gaps are foundational, which are the applied AI layer, and which are the operations layer that turns a demo into a service, so the learning order is not left to whichever tutorial ranks first.
How it works
Researcher versus engineer
The researcher (01:24) works on foundational models: new architectures, training from scratch, papers; the entry ticket is mathematics and usually a graduate degree (01:34). The AI engineer (01:44) starts from existing models and composes them into software: connect data, grant tool access, build memory loops and guardrails, ship a solution that works (01:58). Engine builder versus car builder (02:24). The distinction is about where value is created, not about prestige: most organisations will never train a foundation model and will need many people who can build reliable systems on top of one.
Tier 1: engineering foundations
The prerequisites before any AI-specific skill (03:03):
- Python literacy, and not only for writing. The engineer must read and verify code written by agents and understand the libraries underneath, PyTorch and TensorFlow among them (03:16).
- Git, the command line, Linux. AI tooling and runtimes live on Linux; version control and terminal fluency are assumed (03:50).
- API design and integration. An AI system is a sequence of well-structured API calls between the application, the model and external tools, with response handling and rate limits managed deliberately (04:12).
Tier 2: applied AI skills
- Embeddings and vector search (05:07). Text becomes a high-dimensional vector that captures meaning, so search works by semantic similarity rather than keyword match.
- Retrieval-augmented generation, RAG (05:46). Internal documents and policies are chunked, embedded and stored in a vector database; at query time the relevant passages are injected into the model's context, which grounds the answer and reduces hallucination.
- Agents and tool use (07:29). Unlike a fixed workflow, an agent chooses tools dynamically, calls an API or a database, observes the result and loops until the goal is met. The skill is building agents that are reliable, which is a harder word than "impressive".
Tier 3: deployment and operations
The tier that turns a prototype into a service (07:54):
- Containers and Kubernetes (08:14). Package the agent and its runtime so it deploys the same way on hybrid cloud or bare metal.
- Observability and tracing (08:30). Record which data and which tools an agent went through to reach a decision, so the decision can be explained and trusted.
- Monitoring (08:53). Token cost, latency and security posture, watched continuously rather than at launch.
| Tier | Skills | The question it answers |
|---|---|---|
| 1 Theory | Python, Git/CLI/Linux, APIs | Can you read, run and connect things? |
| 2 Applied AI | embeddings, RAG, agents and tools | Can you make a model useful on your data? |
| 3 Operations | containers, tracing, monitoring | Can you keep it running, explainable and affordable? |
The three use cases employers ask for
RAG knowledge systems for HR, hospitals and internal wikis, answering from the organisation's own documents (09:30); tool-using data-analysis agents that run database queries and produce visualisations (09:48); and AI-assisted deployment automation that shortens build and release cycles (10:03).
Where it lands in Jayverse
- Every repo: grade the tiers honestly. Tier 1 is solid (TypeScript rather than Python, Git, CLI, Cloud Run). Tier 2 is partial: agents and tool use exist in Rabbit's mandate flow and in the Claude Code workflows, but there is no RAG over our own docs yet; the alice repo with its 1,200 markdown pages is the obvious corpus for a first retrieval experiment. Tier 3 is the weakest: tracing of agent decisions is mostly absent.
- Rabbit: an agent that moves money needs Tier 3 first. The video's observability point is the same as the Auditor's rule: a mandate execution must record which data and which tool calls led to the transaction. Add a trace per agent run before adding agent features.
- Tech #62 and the boundary file: this is where the judgment lives. The video's "judgment about architecture" is what the boundary document captures: interface, invariants, allowed tools, evaluation. The two items together say what an AI engineer produces when the code is produced by an agent.
- Eng: the interview vocabulary. Researcher versus engineer, the three tiers, RAG, agent loop, observability: this is the shared language of AI engineering job posts abroad. The next Eng conversation should use it.
Verified and unverified
Verified on 2026-09-19: IBM Technology publishes an explainer video under this title, and the researcher-versus-engineer framing, the three tiers and the three use cases match the summary. Taken from the summary and not independently checked: the exact timestamps and phrasings. The tier structure is the video's, not an industry standard; other roadmaps put data engineering or evaluation as their own tier, and evaluation (how you know an agent is reliable) is the notable gap in this one.
Sources: IBM Technology — Essential Skills for Becoming an AI Engineer: RAG, AI Agents, & More, YouTube · related items: Tech #62 (agentic engineering writes the boundaries), Tech #63 (a model is weights plus an objective), Eng #32 (how do you use AI).
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| foundation model | 기초 모델(대규모로 사전 훈련되어 여러 작업의 바탕이 되는 모델) · 연구자 쪽 용어. "trains foundation models from scratch" |
| frontier model | 프론티어 모델(현재 성능 최전선의 상용 모델) · 오픈소스 모델과 짝으로. "frontier or open-source models that already exist" |
| wire … into | ~를 …에 엮어 넣다(연결해 시스템으로 만들다) · 통합을 말하는 동사구. "wires them into systems that create value" |
| guardrails | 가드레일(모델이 벗어나지 못하게 하는 제약) · 안전·정책 층을 부르는 관용어. "memory loops and guardrails around them" |
| judgment | 판단력(무엇을 왜 쓸지 결정하는 능력) · 이 영상의 핵심 단어. "engineering judgment is the differentiator" |
| differentiator | 차별점(남과 구분되게 하는 요소) · 채용·전략 어휘. "judgment is the differentiator" |
| literacy | 문해력(읽고 이해하는 능력, 여기서는 코드) · "Python literacy" 처럼 언어 이름 뒤에. "Python literacy, and not only for writing" |
| rate limit | 레이트 리밋(단위 시간당 호출 상한) · API 연동의 기본 제약. "rate limits managed deliberately" |
| embedding | 임베딩(의미를 담은 숫자 벡터) · 벡터 검색의 입력. "Embeddings and vector search" |
| semantic similarity | 의미 유사도(뜻이 얼마나 가까운가) · 키워드 일치(keyword match)와 대비. "by semantic similarity rather than keyword match" |
| RAG | Retrieval-Augmented Generation(검색 증강 생성) · 문서를 찾아 컨텍스트에 넣고 답하게 하는 패턴. "Retrieval-augmented generation, RAG" |
| chunking | 청킹(문서를 조각으로 나누기) · RAG 전처리 단계. "chunked, embedded and stored" |
| ground (an answer) | 답을 근거에 묶다 · 환각(hallucination)의 반대말로. "grounds the answer and reduces hallucination" |
| hallucination | 환각(모델이 사실처럼 지어내는 것) · LLM 실패의 표준 용어. "reduces hallucination" |
| agent loop / observe | 에이전트 루프(행동-관찰-판단 반복) · 고정 워크플로와 대비. "observes the result and loops until the goal is met" |
| observability | 관측 가능성(내부 상태를 밖에서 알 수 있는 정도) · 로그·메트릭·트레이스의 상위 개념. "Observability and tracing" |
| tracing | 트레이싱(한 요청이 거친 경로를 기록) · 에이전트 결정 설명의 도구. "which data and which tools an agent went through" |
| bare metal | 베어메탈(가상화 없는 물리 서버) · 하이브리드 클라우드와 짝. "on hybrid cloud or bare metal" |
| security posture | 보안 상태(전체적인 방어 수준) · 운영 모니터링 항목. "Token cost, latency and security posture" |
| syllabus | 강의 계획서 · 로드맵의 비유. "for the same reason a syllabus does" |