About VietAlpha

Our purpose and mission are distinct. The purpose is why VietAlpha exists:

The mission defines what the lab does today. We expect to revise it as the work grows:

Everything we have shipped covers the modern curriculum. The historical texts are next, and we are selecting our first project now.

Digitization comes first because everything else depends on it. Because generic document parsers corrupt math syntax, we design new algorithms specifically for Vietnamese math. Our parser, Cadmus, uses a dual-view normalization-chunking segmenter to isolate math symbols for encoding while preserving 100% of mathematical tokens in the output. We benchmark these pipelines against standard baselines to guarantee data integrity before publishing the results as open datasets. We built Cadmus for the modern curriculum. Adapting it to historical manuscripts, with their distinct scripts and layouts, is our main engineering challenge.

The second part is studying these digitized texts. We analyze them to see how Vietnamese mathematical problems developed and how historical calculations compare with modern lessons. Almost no one else works on this: a few scholars, working outside Vietnam, have studied individual treatises, and the history of Vietnamese mathematics remains a nearly empty field. That is a strange place for a student lab, but it means a small team digitizing these texts builds the foundation for everyone who follows.

The third part is applying these digitized materials to technology. Our first tool is LumiToán, a free public library of the modern curriculum that contains Lumi, an AI tutor fine-tuned on our dataset. More than 10,000 students use LumiToán each month. This usage improves the database: every lesson runs on digitized materials, and student corrections return to the dataset.

The work forms a loop: the record becomes the dataset, the dataset trains the model, the model runs in the classroom, and student corrections return to the record. Every tool we build draws on this record. This makes each new tool cheaper to build and more reliable. This is the difference between VietAlpha and products that simply wrap external models: our model is built on a record we digitized and structured ourselves using custom, benchmarked segmenter algorithms.

The first is survival. The historical collection exists on paper in only a few libraries, and paper does not wait. A digitized text published in an open dataset cannot be lost to fire, flood, or neglect. A studied text cannot be lost to illegibility once we record how to read it. Preservation saves the book. Study saves the mathematics.

The second is that the history of mathematics, as a field, says nothing about Vietnam. Vietnam is absent from the global story of the subject, which scholars wrote from the records they could read. Absence of study is not evidence that nothing happened: a country that ran its administration and its examinations on arithmetic for centuries has a mathematical past. The only way to learn what that past holds is to read it, and reading it starts with digitizing it.

The third is teaching. The surviving texts show how mathematics was explained to Vietnamese learners, in their own terms, centuries before schools adopted foreign textbooks. We expect this material to yield problems, methods, and approaches that belong to the tradition students already stand in. We do not yet know how much of it will transfer to modern lessons; our initial analyses should tell us.

The fourth is technology. A scan is only a picture, and a model cannot learn Vietnamese math from a picture. Analysis turns digitized pages into structured knowledge: problems paired with solutions, methods broken into steps, and terms mapped to meanings. We design new algorithms specifically for Vietnamese math to perform this translation, using benchmarks to verify that our parsing and chunking steps lose no mathematical syntax. Lumi shows that this structured, benchmarked material becomes teaching capability. This applies to the historical collection too: studying the texts is what makes them buildable.

We must state one boundary. A few scholars, working outside Vietnam, have studied individual treatises, and our work covers the modern curriculum. No systematic digitization or study exists, and we are stepping into that opening. An empty field is a strange place for a student lab. But it means a team digitizing these texts builds the foundation for everyone who follows.

We believe Vietnamese students should be able to read their country's mathematical past. That is why everything this work produces (editions, translations, and datasets) will be open.

In Vietnam, premium tutoring for a single exam subject costs 7.2 million VND a month. The average rural income is 4.5 million VND (external source) a month. A month of tutoring costs more than a month's income. This divide widens: the wealthiest families spend ten times more (external source) than the poorest on tutoring. In a system where national exams decide a student's path, exam preparation is priced admission.

Building a free alternative required data. No digital math dataset existed for the Vietnamese curriculum, leaving nothing to train a model on. We built the pieces in order: Cadmus parsed scanned papers to produce the first math reasoning dataset built to MoET standards. From that dataset, we built LumiToán. The work won first prize in the Edtech category of the ICIA National Youth Innovation Award.

Building LumiToán taught us where the scarcity lay. Building models was not the barrier: compute is rentable, and standard language models improve each quarter. The barrier was that existing algorithms failed on Vietnamese math, shredding logical problem units and discarding math markup. We realized we needed both clean data and new algorithms designed for Vietnamese math. The foundation under every tool we build is the record itself, and no one was digitizing, benchmarking, or studying it.

So VietAlpha reorganized as a research lab. Our original mission remains: we began as an answer to the cost of tutoring, and we still provide free, curriculum-aligned math help to students. What changed is the focus. We now design custom segmenters and normalization algorithms to parse Vietnamese math, benchmark their accuracy, and digitize the texts first so every tool can draw from them. Free access is what the work produces; the record and the algorithms to read it are what the work is about.

We publish research and engineering notes as they happen, on the research and news pages. We release our datasets openly, so anyone building for Vietnamese students can start from verified data instead of scanned paper. We treat LumiToán as the first application of our work, not the entire project: it is proof that the loop works, not its conclusion. The historical program is our newest work and the reason we reorganized. It is a part of the record that a product roadmap would not reach.

Three commitments remain unchanged.

CS Inspiration is a network of education programs run by students. VietAlpha is its research arm. The lab receives cloud support from Amazon Web Services and Microsoft Azure. We work with institutional partners in Vietnam and the United States, including the Massachusetts Institute of Technology, the UEH Institute of Innovation, Ho Chi Minh City University of Technology, and Deerfield Academy.

VietAlpha is student-run and small.

Find our research at /research and announcements on News. LumiToán is at lumitoan.com↗ (external site). For everything else, contact harry@csinspiration.org.