The costliest mistake in AI training in Nepal is paying for the wrong kind. Learning to use ChatGPT at work, learning to build AI apps, and doing paid "AI training" tasks are three different things, and each needs a different first step.
The other mistakes come after you pick. Most of them cost months, not money. Here they are, with what to do instead.
What does "AI training" mean in Nepal?
When you search for AI training in Nepal, most news results are about short workshops. They teach office staff, teachers and students to use AI tools well. That is useful, but it will not prepare you for an AI engineering job.
AI engineering is different. You write Python that calls a model, feeds it your own documents, tests its answers and runs it online for real users.
There is also a third meaning. Remote job boards post "AI training" work, where you rate AI answers, write code to grade them, or check Nepali text. That is a job, not a course.
Decide which one you want before you pay anyone. We wrote earlier about why AI and data science skills are in demand in Nepal. The rest of this article is about the second meaning: learning to build.
Do you need Python before you learn AI?
Yes. Skipping Python to jump straight to prompts is the mistake we would warn you about first. Every AI engineering tool you will use, from scikit-learn to LangChain, expects you to read and write Python.
Even the free Hugging Face LLM course says it "requires a good knowledge of Python". Without it, you copy code you cannot fix.
Classic machine learning matters too. Some problems, like sorting short texts into a few groups, are solved faster and cheaper by a simple model than by an LLM. Learn when that is true before you reach for the biggest tool.
Do you need a gaming laptop or a GPU for AI?
No. Do not buy new hardware to start. A normal laptop with 8GB of RAM is enough for the first months.
For heavy steps, use Google Colab. It is free, though its free GPU access is "heavily restricted" and a notebook can run for at most 12 hours. That is enough for learning.
To run a model on your own machine, install Ollama and type ollama run gemma4. On a laptop with 8GB, pick a small model. It will be slow, but once it is downloaded it runs without internet, which helps on days when the connection drops.
Should you pay for ChatGPT or Claude API credit to start?
Not yet. Google AI Studio has a free tier for the Gemini API, and Groq has a free plan too. You can learn prompts, structured outputs and streaming without a bank dollar card.
There is one catch. Google says content sent on the free tier is used to improve its products. So never paste company data, client files or anyone's personal details into a free-tier app.
Should you start with LangChain or agents?
Start without them. Call the model directly first, in a few lines of Python. Then you see exactly what goes in and what comes out.
Anthropic's guide on building effective agents gives the same advice. It says frameworks "often create extra layers of abstraction that can obscure the underlying prompts". It also recommends the simplest solution that works, and adding agents only when you need them.
Once you can build a small agent loop by hand, a framework saves you time. Before that, it hides your bugs.
Is a working demo enough?
No. A chatbot that answers your five test questions well can still fail on the sixth. Learn to measure answers against a fixed set of questions, and track what each request costs.
Learn safety early too. Prompt injection, where text tricks the model into ignoring your instructions, is first on the OWASP Top 10 for LLM applications. Any app that reads emails, web pages or uploaded files is open to it.
How long is AI engineering training at Saarathi?
The AI Engineering and Machine Learning course at Saarathi Academy in Old Baneshwor, Kathmandu, runs for 12 weeks, in the order this article argues for.
Weeks 1 to 5 cover Python from zero, machine learning, NLP and deep learning before any LLM work. Week 4 includes a text classifier with a Nepali run, and a session on when a simple model beats an LLM. Class runs on free tiers, so you do not have to pay for API credit. In Week 9 you build the agent loop by hand before using LangChain. Week 10 covers evaluation and prompt injection.
The course is fast. Plan for 2 to 3 hours of practice every day on top of class. A laptop with 16GB of RAM is recommended, and 8GB works with small local models and free Colab. Each batch has up to 10 students in the classroom, and mentors still do the job they teach.
What to try tonight
Install Ollama, run ollama run gemma4, and ask it to explain one Python error you got this week. Then open Google AI Studio, make a free key, and call Gemini from a Python file. If that second step feels hard, spend a few weeks on Python basics first. Now you know where you stand.
The week-by-week plan is on our course page, and the full syllabus is public before you pay anything.
Top comments (0)