The first four weeks to become an AI engineer in Nepal, before any LLM

To become an AI engineer in Nepal, spend your first four weeks on Python, data and simple machine learning, not on ChatGPT APIs. AI engineer job posts in Kathmandu ask for strong Python and SQL first. The habit of testing a model on data it has never seen is the same habit you later use to test chatbots.

So LLMs wait until week 6. Below is the order we would follow for the first four weeks, and the reason for each step.

What do AI engineer jobs in Nepal ask for first?

Read the job posts before you plan your learning. A mid-level AI engineering post on Kumari Job, now closed, put "Strong Python skills" at the top of its required list. SQL and backend work came next. LangChain and vector databases sat lower, under preferred skills.

A 2025 AI software engineer post on merojob asked for pandas, NumPy and scikit-learn as required skills. Kumari Job's own career guide says entry-level hiring looks for practical Python and SQL.

The pattern is clear. Employers expect the LLM tools to sit on top of solid basics.

Week 1: Python basics you write yourself

Learn values, variables, strings and f-strings. Then learn if and else, for and while loops, lists and dictionaries. Lists start counting at zero, and that is the most common beginner bug.

F-strings matter more than they look. Almost every prompt and log line you write later is an f-string.

You can turn on an AI autocomplete like Cursor from day one. But read every line it writes, and type the small programs yourself first.

If you have never coded, do not start with the official Python tutorial. Its first page says it is written for programmers new to Python, not for people new to programming. Python for Everybody is free and fits a first-time coder better.

Week 2: the Python habits AI code depends on

This week is functions, type hints, try and except, and reading and writing JSON files. Then take a first look at numpy and pandas.

Each topic has a direct use later. LLM APIs send back JSON. API calls fail when you hit a rate limit, so your code must catch the error and retry. Type hints let your editor catch mistakes before you run anything.

Skip this week and your first chatbot breaks on its first bad reply.

Week 3: data, SQL and your first model

Load data into a pandas DataFrame, then filter it, group it and merge it. Write SQL with SELECT, WHERE, JOIN and GROUP BY, and read the result into pandas.

Then train your first models with scikit-learn. Linear regression predicts a number. Logistic regression predicts yes or no.

The key lesson of the week is the split. You keep some data hidden, train on the rest, and judge the model only on the hidden part. Learn why accuracy lies when one answer is rare. If only 1 case in 100 is fraud, a model that always says "not fraud" scores 99 percent and catches nothing.

Week 4: testing a model honestly

Learn overfitting. That is a model that memorised the training data and fails on anything new. Then learn cross-validation, which tests the model on several different splits instead of one lucky one.

Add decision trees and random forests. Then build a text classifier the classic way, with TF-IDF and logistic regression, to sort messages by topic. Try it on some Nepali text and watch where simple word splitting breaks.

End the week by comparing this small model with the cost of calling an LLM for the same job. For a simple sorting task, the small model is often cheaper and faster. Employers test that judgement.

Why not start with ChatGPT and LangChain?

Calling an LLM API takes an afternoon. Knowing whether its answers are right takes weeks of habit.

When you build RAG later (an AI that answers from your own documents), you need a test set of questions with known answers. You need to score it on data it did not see. You need to spot when the test leaks the answer. Those are the week 3 and week 4 skills, used on a chatbot.

Skip them and you build demos that work once, then fail in front of a real user.

Can you learn this on the laptop you have?

Yes. None of the first four weeks needs a graphics card. scikit-learn and pandas run on an ordinary laptop with the small datasets beginners use.

For heavier steps later, Google Colab runs Python in your browser for free. Its free GPU access is limited and not guaranteed, and an idle session times out. So save your work often, especially when your internet or power drops.

What to try tonight

Open the scikit-learn getting started page. Copy its pipeline example into a free Colab notebook: make_pipeline(StandardScaler(), LogisticRegression()) on the Iris data, split with train_test_split. Run it and read the score. Then run cross_validate on the same model and see how the score moves across five splits.

If that takes you one evening and you understand each line, you are ready for week 3 of this plan.

At Saarathi Academy in Old Baneshwor, Kathmandu, the AI Engineering and Machine Learning course follows this order. Python is taught from zero in weeks 1 and 2, machine learning runs to week 5, and LLMs start in week 6 of 12. The full week-by-week syllabus is public before anyone pays: read the syllabus, or see the course page.