The students who could benefit most from an AI tutor may be the ones least able to use one.
Most of today’s generative AI tools assume a user has reliable internet, enough data and access to reasonably capable hardware. Filipino developer Luis Buenaventura is experimenting with the opposite premise: what if the tutor lived entirely on an inexpensive phone?
His project, Hiraia, is an open-source AI science tutor designed for Filipino elementary and junior high school students. After its initial files are installed, it can operate without an internet connection, with lessons and conversations available in Tagalog, Bisaya and English.
And that offline design addresses a very Philippine problem.
AI without the broadband bill
Only 48.8 percent of Philippine households had internet access at home in 2024, according to the Philippine Statistics Authority. Among households without access, the most commonly cited obstacle was the high cost of an internet subscription, followed by the cost of equipment.
Hiraia attempts to remove the first of those recurring costs.
Instead of sending every question to a cloud data center, its AI model runs directly on an Android device. The target is relatively modest hardware: Android 12 or newer with 6GB of RAM.
Getting several gigabytes of AI files onto a phone still presents an obvious problem for a student without broadband. Hiraia’s answer is peer-to-peer distribution: after one device obtains the necessary files, other devices on the same local network can receive them without each student separately downloading the entire package from the internet.
Once installed, the system requires no account and is designed so that what a child types does not leave the device.
That makes offline AI not merely a connectivity feature, but a privacy feature as well.
Teaching AI to speak like its students
There is another obstacle that faster internet alone cannot solve: language.
Research published with FilBench, a benchmark specifically created to test large language models on Filipino, Tagalog and Cebuano tasks, found significant weaknesses even among sophisticated models, particularly in areas including reading comprehension and translation.
Hiraia is therefore being trained specifically around the languages its intended students use.
Its underlying small language model has been further trained on Filipino and Bisaya material, while its educational content is aligned with the Department of Education’s MATATAG science competencies. The project currently describes a library containing 50,000 science facts, 30,000 illustrations and 20,000 mini-quizzes.
Rather than allowing the AI unlimited freedom to invent educational material, the system retrieves information from a prepared fact bank and uses pre-generated learning materials — an attempt to reduce the hallucinations that remain a fundamental problem with generative AI.
A tutor, not a teacher replacement
Buenaventura says the ambition is not to replace classroom teachers.
“My hope is that I can roll this out to public schools across the country as a supplemental teaching tool — not to replace teachers, but to reinforce classroom learning with some personal tutoring at home,” he said.
That distinction happens to align with where Philippine education policy is heading.
DepEd’s Basic Education AI Framework, issued this year, requires AI in schools to remain human-centered and pedagogically appropriate, with teachers and learners retaining control. It also specifically calls for AI access to include poorer students and those in geographically isolated and disadvantaged areas.
The department has simultaneously begun building its own AI capabilities, including an Education Center for AI Research and other AI-assisted education initiatives.
Hiraia, however, remains an experiment rather than a proven educational intervention.
The software is still in early alpha, and Buenaventura is seeking an academic institution with which to conduct controlled pilot trials.
That testing will matter. Being able to run AI on a cheap phone is an engineering achievement; demonstrating that children actually learn better with it is a separate question.
But Hiraia is already testing an intriguing reversal of the usual technology equation.
For years, bringing sophisticated educational software to underserved students largely meant waiting for better devices and better connectivity.
Hiraia is asking whether the technology can instead be redesigned around the devices, languages and connectivity students already have.