skip to content
Navid Mesbah
فا
← Back

Persian-First: Building for the People Who Use It

/ 9 min read

The Translation Trap

Most AI products are built in English and then “localized.” Translate the interface strings, flip the layout to right-to-left, check that nothing overflows, and ship. A checkbox on a launch plan: Persian support, done.

For Persian speakers, the result is familiar. A product that technically speaks your language but clearly wasn’t thinking in it.

The numbers break: sometimes Persian digits, sometimes Latin, sometimes both in the same line. Mixed Persian and English text jumps around as the cursor moves, so that editing a sentence with a single English word becomes a small puzzle. Punctuation lands on the wrong side. Fonts are an afterthought, often a default that renders Persian letters thin, cramped, or with the wrong joining forms. Examples are culturally foreign: dollars, miles, Thanksgiving, baseball. And the model underneath often reasons in English and translates at the end, which shows in strange phrasing, lost nuance, and an unmistakable feeling of reading a translation.

None of these problems is individually fatal. Together, they send a clear message: this product was made for someone else, and you are a guest in it.

I’ve been building a Persian-first engine for math and science explainers, and the main lesson is that “Persian-first” is a design decision, not a translation task. It has to be made at the beginning, and it affects everything that follows.

Why I Care About This

This is personal for me.

For years I worked with students as a university teaching assistant and as a math, tech, and computer science coach, mostly with Persian speakers. Many of them were bright, curious, and capable of deep understanding. But much of the best educational material in the world, the beautiful interactive explainers, the carefully produced videos, the thoughtful textbooks, was in English. Students with strong English could access it. Others couldn’t, or could only access it through a fog of translation that made difficult ideas even harder.

I watched talented students conclude that serious knowledge lived somewhere else, in another language, for other people. That conclusion is wrong, and it’s corrosive. It makes people feel like visitors in the world of ideas rather than citizens.

Persian has a deep scientific and literary history. Some of the most important mathematicians and scientists in history wrote in Persian or in the Persian-speaking world. The language is perfectly capable of carrying precise, beautiful, rigorous thought. What it lacks is modern tools built with the same care as the English ones.

AI could make this gap much worse or much better. If AI products continue to treat Persian as an afterthought, the gap will grow, because AI-generated educational content will flood the world in English first. If we build Persian-first tools, AI could finally make high-quality, interactive, personalized learning available in Persian at the same level as anywhere else.

What Persian-First Means in Practice

Typography first. Persian text with embedded math, numbers, and Latin terms is genuinely hard to lay out. Equations read left-to-right inside right-to-left sentences. Variable names are usually Latin letters. Numbers might be Persian or Latin digits depending on context. Units, chemical formulas, and code snippets all have their own direction. A single sentence in a physics explainer might switch direction four or five times.

If this isn’t solved at the foundation, nothing built on top looks right. So we treated bidirectional layout as a core engineering problem from day one, not a bug to fix later. We tested with real mixed content, built a large set of tricky cases, and made sure every component, from text rendering to equation display to animated labels, handles direction correctly.

Font choice turned out to matter more than I expected. Persian script has a long calligraphic tradition, and readers are sensitive to the quality of letterforms. A font that renders Persian well signals respect; a poor one signals neglect. We chose and tuned fonts carefully, and paired them with mathematical typography so that Persian text and equations feel like they belong on the same page.

Language that sounds like a teacher, not a translator. There’s a natural way a good Persian teacher explains a derivative or a pendulum. It’s different from a translated English textbook. The rhythm of the sentences is different. The rhetorical moves are different. The examples are different. Even the jokes are different.

A translated explanation might say something grammatically correct but stiff, the way a formal document sounds. A good Persian teacher speaks more warmly, uses familiar expressions, asks questions of the student, and builds intuition through images that resonate culturally.

We write prompts, examples, and evaluation criteria in Persian from the start. The model is asked to think about how a skilled Persian-speaking teacher would explain the idea, not to produce an English explanation and translate it. This makes a noticeable difference, and native speakers notice it immediately.

Terminology with care. Scientific Persian has a complex vocabulary situation. Some concepts have well-established Persian terms. Some are usually referred to with borrowed words from French or English. Some have multiple competing terms. Some have official terms that nobody actually uses.

Students need to learn the terms they’ll actually encounter in their schools, exams, and future reading. So we built terminology guidelines grounded in the actual Persian educational system, and we show alternative terms where it helps understanding. This is unglamorous work, but getting it wrong confuses students and erodes trust.

Evaluation in Persian. If you only evaluate in English, you’ll ship something that’s great in English and mediocre in Persian, and you won’t even know. Aggregate quality metrics hide language-specific failures.

Our evals are written and reviewed by Persian speakers. They check not only correctness but naturalness, clarity, terminology, and cultural fit. Some failures only a native speaker would catch: an expression that’s technically correct but sounds odd, an example that would confuse a student from a particular background, a tone that feels condescending.

Context that belongs to the learner. The examples in an explainer should feel like they come from the learner’s world. Probability can be explained with backgammon dice, a game deeply woven into Persian culture. Geometry can draw on the patterns of Persian tilework and architecture, which are themselves extraordinary works of mathematics. Rates and ratios can use the familiar experience of bargaining in a bazaar. Physics can use the trajectory of a ball in a street football game, the cooling of a cup of tea, the way snow sits on the mountains around a city.

When a student sees their own world in an explanation, something shifts. The math stops being a foreign object and becomes a way of seeing things they already know.

The Technical Reality

I want to be honest about how hard this is technically.

Large language models are trained on data that is overwhelmingly English. Their Persian is often good, sometimes very good, but less reliable. They’re more likely to make factual errors in Persian, more likely to produce awkward phrasing, and more likely to slip into English-style reasoning. Their knowledge of Persian educational conventions and terminology is thinner.

We work around this in several ways. We provide rich Persian context and examples. We use verification steps that check mathematical content independently of the language it’s expressed in, so that a correct explanation in Persian isn’t penalized and an incorrect one isn’t rescued by fluent prose. We generate visual and interactive content from computation rather than from language, so that a graph is always correct even if a sentence about it needs polishing. And we invest heavily in Persian-specific evaluation to catch what the models miss.

It’s more work than building in English. Considerably more. But every bit of that extra work is visible to the people using the product, and they deserve it.

Why It Matters Beyond Persian

Persian has tens of millions of native speakers and a deep literary and scientific history, but it’s underrepresented in AI training data and product design. The same is true for many languages: Bengali, Swahili, Amharic, Tagalog, Pashto, Kurdish, and hundreds more. Collectively, the speakers of these languages are a large part of humanity.

When products treat these languages as an afterthought, speakers learn a quiet lesson: serious knowledge lives in another language. Over time, this can erode the languages themselves. Young people choose to study and work in English because that’s where the tools and resources are, and the scientific and technical vocabulary of their own language atrophies.

I don’t want students to learn that lesson. A student in Shiraz, Tabriz, Mashhad, Herat, or Dushanbe should be able to explore calculus in their own language at the same quality as anyone else in the world. Not as a charitable afterthought, but because their language and their curiosity deserve it.

And there’s an opportunity here, not just a responsibility. Persian-speaking communities are large, educated, and underserved by high-quality digital tools. Building well for them is not only right, it’s also an open space where careful work can make a real difference.

Advice for Other Builders

If you’re building for a non-English audience, here’s what I’ve learned.

Put a native speaker in the loop from day one, not at the end. Not just a translator, but someone who understands the domain and the culture. They’ll catch problems that no amount of translation review will reveal.

Test with real mixed-language input. Your users will mix languages in ways you don’t expect. If your product breaks when someone types a Persian sentence with an English word in the middle, it’s broken for most of your users.

Treat layout and typography as core engineering. Right-to-left and bidirectional text are not cosmetic problems. They affect every component. Solve them at the foundation.

Measure quality in the target language directly. Don’t assume English quality transfers. It doesn’t, at least not fully.

Ground your work in the real context of your users: their education system, their culture, their daily life. Generic localization produces generic products.

It’s more work up front. It’s also the only way the product feels like it was made for the people using it. And when it does, they notice, and they tell you. Some of the most meaningful feedback I’ve received in my career has been from people who said, simply, that for once it felt like something was made for them.