A Separate Vacancy: Fullstack Python + React + TypeScript Who Already Shipped LLM to Users

The gap on the board is not “someone who also tried GPT.” It is a fullstack seat: Python on the API, React and TypeScript on the operator screen, and an LLM feature that already met real users — abort in the middle of a stream, a 40-page paste, a double tap on send. That bar is the point of the posting.

I already wrote the generic fullstack role, and the Python AI-engineer demand that names LangGraph and RAG in production. This piece is the other vacancy. They are not a softer version of each other.

1. Why this posting exists

A lot of products cannot hire an AI engineer, a backend, and a frontend for one feature. They need one person who already crossed the product boundary with a model. The JD is a budget decision as much as a skills list.

• The LLM work is a surface in an existing app, not a retrieval platform. The user already has a screen. The model has to live on it.
• “Shipped to real users” is the filter that kills the hackathon. Localhost does not count. A café demo does not count.
• Python plus React plus TypeScript is named because the stack is already split that way — FastAPI or Django on one side, a typed UI on the other — and they will not fund a rewrite to Next-only or Python-only.

2. What “solid on both sides” means in the posting

It does not mean “I can read the other repo.” It means the person has shipped a typed contract between the two runtimes and did not let them drift on a Friday.

• Python: the API, auth, jobs, the adapter to the model, secrets that never leave the server.
• React + TypeScript: abort, empty and degraded states, citations as UI, a confirm card for a tool — not a markdown dump from the model.
• One schema: Pydantic on the server, TypeScript from OpenAPI or generated Zod. Two hand-written interfaces that drift are how this hire fails in month two.

3. What “already shipped to real users” filters out

Interview for the scars, not the screenshot. Users will cancel, paste a book, and tap send twice. If the candidate cannot name those three, they have not shipped.

• Auth on the stream, not only on the page. A leaked EventSource URL is a leaked corpus.
• Streaming as events a UI can tell apart: token, citation, tool_propose, error, done. Concatenated strings cannot show a source or a confirm.
• A named error when the provider is down. I wrote the ChatGPT-outage playbook. The product needs a string, not a blank bubble.
• A feature flag. Shipped means they could turn the model path off without rolling back the site.

4. Who posts this — and who should not

This vacancy is a fit when the product already has a UI, the backend is already Python or will stay Python, and the LLM feature is one slice of the app. It is a mismatch when the job is a corpus platform with evals and a graph that must refuse.

• Seed and small product teams, internal tools, agencies that ship a copilot into an existing React app.
• Not a substitute for the Python-first AI engineer when the pain is stale retrieval, faithfulness, and a graph in production. That demand is the other article.

Conclusion: two seats, two prices

Fullstack here means the Python loop and the React screen share a schema, a stream, and a confirm — and that this already happened in front of users. A notebook with a React wrapper is still a notebook. Write this vacancy as its own seat. If the same JD also names LangGraph, RAG pipelines, and an LLM system kept in production, you are shopping two seniors. Price it that way, or split the posting.