The Market Wants a Python AI Engineer With LangGraph — Not an “AI Person”

Vacancies stopped saying “we need AI.” They now name LangChain or LangGraph, a RAG pipeline, and an LLM system the person has kept running in production. That sentence is a filter. It is not a nicer title for a prompt freelancer, and it is not the same seat as a fullstack Python-and-React hire with a model in the product.

I already wrote what an AI engineer does in general. This piece is the demand that actually shows up in 2026 postings: Python-first, hands-on LangChain or LangGraph, RAG as a pipeline, and production — not a demo. The other vacancy on the market is a separate article.

1. Why this posting appeared

Companies shipped a chat box. Quality did not ship with it. The corpus changed on Monday; Tuesday’s answer still cited last month. That is when the JD stopped saying “AI” and started saying RAG, LangGraph, and production.

• Internal search, support deflection, policy Q&A: the product is retrieval that stays current, not a personality.
• A contractor demoed a LangChain chain. Week three broke. The next posting asks who kept a system running, not who imported a library.
• Model labs and OWASP made “unsupervised agent” a first-class risk. Buyers want a person who can bound the graph, not a larger prompt.

2. What each keyword in the JD actually buys

Read the posting as a shopping list. If you treat LangGraph as a synonym for “knows ChatGPT,” you will interview the wrong people.

• Python-first: the turn lives in Python — FastAPI or a worker, typed state, ingest as a job. A notebook pasted into Node is not this seat.
• LangChain or LangGraph hands-on: they want the graph — retrieve, grade, rewrite, refuse — not `import langchain`. Empty retrieval is a node they can log, not a surprise in the prompt.
• RAG pipelines: ingest, chunking by document type, hybrid search, citations as ids. Upload-to-a-vector-db is not a pipeline. I wrote why the vector index is a derived index, not the system of record.
• An LLM system they kept running in production: evals that can fail a deploy, a stale-index story, a fallback when the provider is 500, a kill switch. Fluency in a café is not this bar.

3. Who is buying this hire

Not a marketing team that wants “ChatGPT on the homepage.” The budget sits where a wrong answer has a cost: tickets, compliance, an internal wiki people already distrust.

• Product companies with a private corpus: policies, SKUs, contracts, runbooks.
• Support and ops teams that already measured first-response time and want retrieval, not another empty inbox.
• Teams that tried an agent with write access and now want a graph with a refuse node. I wrote why the till stays with a human.

4. What a mis-hire looks like

The market is loud. The wrong interview loop is louder. These three profiles fail this JD even when the CV says AI.

• An ML researcher who trained nets and never owned ingest, evals, or a Tuesday reindex.
• A frontend engineer who wrapped OpenAI in a chat UI. That is a product surface. This vacancy is the loop behind it.
• A tutorial: one retrieve, one generate, a screenshot. The posting already had that. It is hiring because it died in week three.

5. This is not the fullstack LLM vacancy

A second posting on the same market wants Python plus React plus TypeScript, and an LLM feature that real users already opened. That hire is judged on the contract between the API and the screen. This hire is judged on the graph, the corpus, and whether Tuesday still answers from Monday’s file. Merge the two JDs and you ask one body for two seniors.

Conclusion: buy the loop, not the label

If the JD names LangGraph, a RAG pipeline, and production, the market is buying a Python systems seat around the model. A chain with a chat widget is week one. The demand is for the person who still had an answer after week three. Keep this vacancy off the fullstack LLM posting. They are two seats.