A junior AI engineer in Nepal spends most of the week checking answers, cleaning data and fixing the code around a model. Very little of the week goes into training new models from scratch.

That surprises many people who search "AI engineer Nepal". Most career guides say the job is to "train AI models and improve them". That is true, but it skips the parts that fill the actual hours. This article walks through those parts, so you can practise the right things before you apply.

What do AI engineer job posts in Nepal ask for?

Read a few real posts and a pattern shows. One AI automation post on a large Kathmandu fintech career page asks for Python, LLM APIs from OpenAI, Anthropic and Google, and at least one framework such as LangChain, LlamaIndex or LangGraph. It also asks for vector databases, REST APIs and webhooks.

The same post asks you to "monitor and improve" reliability, latency and cost. Latency means how long a user waits for an answer. An AI engineer post on the same career page asks for pandas, NumPy, SQL and Git, and for experiments that are documented "clearly and reproducibly".

Notice what is missing. No post asks you to invent a new model. They ask you to connect existing models to real data and keep the result working.

What does a normal week look like, task by task?

The exact week depends on the company. But the same five kinds of work come up again and again in these posts.

Checking whether the answers are right

This is the biggest part of the job. A chatbot that answers from company documents can sound confident and still be wrong. So you keep a list of real questions with the correct answers, and you run the bot against it after every change.

Tools help here. RAGAS has a faithfulness score from 0 to 1. It checks each claim in an answer against the documents the bot found. If one of two claims has no support, the score is 0.5. A low score tells you where to look.

Cleaning data before it reaches the model

Documents arrive as scanned PDFs, Excel sheets and mixed Nepali and English text. A model can only answer from what it can read. So a lot of the week is splitting files into clean pieces, removing repeated text and fixing broken characters. Nobody posts about this part. It still decides whether the bot works.

Watching cost and speed

Every call to a hosted model costs money and time. A small change to a prompt can raise both. So juniors are often asked to measure the cost of one request and the wait for one answer, and to report both with each change. A slower but slightly better answer is not always the right trade.

Guarding against bad input

Users type strange things, and some do it on purpose. The OWASP page on prompt injection describes a web page with hidden instructions. A model asked to summarise that page follows those instructions and can leak the private chat. If your bot reads documents or web pages, you test for this before users find it.

Writing it down

Many teams in Nepal work with clients abroad, so updates are written in English. You note what you changed, what the test scores were before and after, and what broke. This feels slow. It is also what lets a senior trust your work without checking every line.

What are the boring parts nobody tells you?

The boring parts are the job. Writing a test set of real questions by hand takes hours. Fixing one PDF that splits badly is slow, careful work. Rerunning the same tests after each small change is repetitive.

Our view is simple. If you enjoy finding out why an answer was wrong, you will like this work. If you only want to watch a demo work once, you will find it hard.

How does this fit a learner in Nepal?

You do not need a graphics card for most of this work. Checking answers, cleaning data and measuring cost all run on an ordinary laptop. Hosted models do the heavy thinking. Free tiers from providers such as Google AI Studio are enough to practise.

Where jobs are posted matters too. Look at company career pages, merojob, KumariJob, JobAxle and LinkedIn, plus remote listings. Read ten posts before you pick what to learn next. The skills they repeat are your plan.

What can you try tonight?

Get a free Gemini API key from Google AI Studio. Keep it in an environment variable, never in your code or on GitHub, as the page warns. Then write 20 questions about one document you know well, such as a college notice, with the correct answer for each.

Ask the model each question with the document pasted in. Mark every answer right or wrong yourself. Then change one line of your prompt and run all 20 again. What you just did is the core loop of the job, done by hand.

How do we teach this at Saarathi Academy?

At Saarathi Academy in Old Baneshwor, Kathmandu, the 12-week AI Engineering and Machine Learning course builds this loop on purpose. In Week 6 you check AI output by hand against your own test set. In Week 10 you add LLM-as-judge, RAGAS and prompt injection defences from the OWASP LLM Top 10. In Week 12 your project goes live with a real cost per request in its README.

Class runs on free tiers, and 8GB of RAM works with free Google Colab for the heavy steps. Batches have up to 10 students in the classroom, and mentors still do the job they teach.

See the course page and the full syllabus.