AI Engineer in 9 Months: 7 Skills in This Exact Order

Nine months, seven skills, one order. This is the path I'd follow again from zero: foundation first, then the LLM layer, then the production layer. One project per step, and no skipping ahead — each step assumes the one before it.

1. Python — enough to read and ship

Start with Python at the level where you can read a script, understand what it does, and write your own that runs from the terminal. That's the bar. The LeetCode grind is a different skill and it doesn't map to this work, so skip it for now — you can come back to it later if you want it.

Official: python.org

bash
python --version

Project: a small script that reads files, hits an API, or cleans a dataset.

2. APIs — every model call is one HTTP request

Strip away the SDK and every model call is one HTTP request: a URL, an auth header, a JSON body, a JSON response. Once you see it that way, debugging gets easier — you're looking at status codes and payloads, not magic. Read the API reference for whichever provider you use before you read a tutorial about it.

Official: OpenAI API docs

bash
pip install openai

Project: call a model from a script, print the raw response, then handle one error case.

3. Prompting you actually test

Prompting is not a bag of tricks. It's a small engineering loop: write a prompt, run it against a few inputs, change one thing, run it again. Keep the prompts that survive that loop and throw away the ones you only liked the first time.

Official: Anthropic docs

Project: a prompt file plus a script that runs it on five inputs and prints the outputs side by side.

4. RAG — chunk, embed, retrieve, cite

RAG is: take your own docs, split them into chunks, embed the chunks, store the vectors, then at query time retrieve the three chunks that matter and answer with citations. Retrieve three, not thirty. Citations are the part that makes it usable, because the answer can be checked.

This is most AI jobs right now.

Official: LangChain · Chroma

bash
pip install langchain chromadb

Project: RAG over your own notes or a docs folder, answering with the source chunk attached.

5. Evals — a test set that scores every change

Without evals you're guessing. Build a small test set of inputs with expected outputs or a scoring rule, then run it on every change to your prompt, model, or retrieval step. The score tells you whether the change helped or just felt better.

Official: Ragas

Project: twenty test cases for the RAG app from step 4, plus a script that prints the score.

6. Agents — tools and error recovery

An agent is a loop: the model picks a tool, you run it, you feed the result back. The hard part is not the happy path — it's recovering when a tool fails, returns garbage, or loops forever. Build for the failure case first, and give the model a way to stop.

Official: LangChain agents

Project: an agent with two or three tools and an explicit retry and fallback path.

7. Deployment — one endpoint, logs, Docker

Ship it: wrap the app in one endpoint, add logging you can actually read, and containerise it with Docker so it runs the same way on your machine and on a server. Then run it and watch the logs for a day.

Official: FastAPI · Docker

bash
docker build -t my-app .

Project: the step-6 agent behind a POST endpoint, containerised, with request logging.

Quick recap

• Python — read and ship scripts, skip the LeetCode grind
• APIs — a model call is one HTTP request
• Prompting you actually test
• RAG — chunk, embed, retrieve three, answer with citations
• Evals — a test set that scores every change
• Agents — tools that recover from errors
• Deployment — one endpoint, logs, Docker
• One project per step, in this order