Learning to code in the language you think in

Here is a situation that happens in every classroom in India and almost never gets named.

A student is reading about recursion. The concept is hard — recursion is genuinely hard, it defeats people with degrees. But the explanation in front of them is also in their third language. When they get stuck, they cannot tell which thing they are stuck on. Is this idea difficult, or did I misread the sentence?

Most people, most of the time, conclude that the idea is beyond them. Some of them are right. Many of them just needed the sentence in Telugu.

Two problems wearing one coat

The confusion between language difficulty and concept difficulty is expensive, and it compounds. A learner who misattributes one to the other draws a conclusion about their own ability. That conclusion is sticky, it arrives early, and it is often wrong.

Our AI mentor explains any lesson in 22 Indian languages. The purpose is to separate those two problems. Read the explanation in the language you think in; if the concept is still hard, now you know it is the concept, and that is a solvable and normal problem. If it suddenly makes sense, you were never confused about recursion.

That is the entire argument. It is narrower than "learn to code in your mother tongue" and we think the narrowness is the point.

What we are not saying

We are not saying English is optional in this industry. It is not, and telling you otherwise would set you up badly.

The documentation is in English. Stack Overflow is in English. Your error messages are in English, your code review will be in English, and if you work with a client in Bangalore or Berlin the meeting is in English. Any course that lets you avoid it is deferring a problem to the point where it costs the most — your first job.

So the keywords stay in English. The code stays in English. The error you paste in is in English and the fixed code that comes back is in English. What changes is the explanation around them.

In practice a Marathi explanation of a Python error reads mostly like Marathi with IndexError, list and range sitting inside it in English, exactly as a Marathi-speaking senior developer would explain it to a colleague at the next desk. That is not a compromise; that is how bilingual technical people actually talk. We are matching the register that already exists rather than inventing a purified one.

The effect is that your English improves in the place it matters — reading errors, reading docs, naming things — while your understanding is not held hostage to it.

Which languages, and why 22

Assamese, Bengali, Bodo, Dogri, Gujarati, Hindi, Kannada, Kashmiri, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, Sindhi, Tamil, Telugu, Urdu.

That is the Eighth Schedule of the Constitution. We used it rather than picking the largest handful by speaker count, because the moment you start choosing which languages are worth supporting you have to defend the cut-off, and there is no honest place to draw it. The schedule is a list someone else already argued about, and it includes languages a commercial calculation would have dropped.

Some of these are better served by current language models than others. Hindi, Bengali, Tamil, Telugu and Marathi are strong. Bodo, Santali, Dogri and Kashmiri have far less training data behind them and the explanations are correspondingly rougher. We would rather offer a rough Santali explanation than none, and we would rather tell you which is which than let you discover it.

Where the AI stops and a person starts

Some limits, stated here rather than in a footnote:

There is a credit quota. AI explanations are not unlimited. Every learner gets a monthly allowance, shown in the portal, and heavy use runs it down. We would rather show you the number than throttle you invisibly and let you wonder why answers got slower.

The human route is always free. This matters more than the quota. Send the exact code and the exact error to your mentor from the same screen, in any language you share with them, with no credit cost and no limit. If the AI is the fast path, the mentor is the reliable one, and we never wanted a design where running out of credits means running out of help.

The AI is sometimes wrong. All of them are. It is very good at "why did this fail", good at "explain this concept another way", and least reliable when asked to design something from scratch. Treat a confident answer about your specific broken code as a strong hypothesis to check, not a fact — which, usefully, is also the correct posture for using these tools in a job.

Why this ends up mattering for employability

There is a version of this feature that is a marketing bullet: 22 languages! That version would not be worth building.

The reason it is worth building is what it does to the drop-off. People leave technical courses in the first six weeks, and when they leave they generally say some form of "it was not for me". A meaningful share of those people were not defeated by programming. They were defeated by reading about programming in a language they were still translating in their head, at the exact moment when everything is unfamiliar and one more source of friction is enough.

Get someone past week six with their confidence intact and the picture changes completely. That is the whole intervention.

You can try the AI mentor on a free demo — bring a piece of code that is broken, in whichever language you would rather have it explained in. Or read more about how the AI mentor works.