Portfolio / Case study

Case study · Education · Non-profit

A free school for every child,
built and written with AI.

SuperEscuela is a non-profit project from Costa Rica: free education for Spanish-speaking children, with no ads, no paid plan and no data sold, that works offline on the oldest phone. A small team could only build it one way: an AI pair programmer for the system, and an AI pipeline for the content.

Education · AI content · Custom software

superescuela.com →

Topics

733

Published, across three age levels

Aligned to the MEP plan of Costa Rica and SuperEscuela’s own path

Lessons

2,786

Written by an AI pipeline

With 13,705 practice questions and unlimited attempts

AI cost

2–4¢

Per lesson, with the higher-quality model now in use

A whole school’s lessons for roughly $60–130 in AI cost

Code

74 of 97

Code changes co-written with an AI coding assistant

Every one accepted by an engineer

Systems

3

Built and running: a student app, a teacher records system and an AI content engine

Plus three public websites

Price

$0

For every student and teacher

No ads, no premium plan, no data sold

Figures from superescuela.org, the project’s own database and its code history, as of 5 October 2026. The AI cost is estimated from the volume of generated content at the model’s list price. SuperEscuela is in public alpha, and its non-profit association is being formally constituted.

How it was built

Five problems,
each with its answer.

733

A whole school, written by AI

The challenge

  • Three age levels of content, aligned to a national curriculum, for a team of a few people
  • Lessons must be short, playful and correct, for children
  • Everything has to work offline, on old phones

What we built

  • A four-stage pipeline: AI plans the topic tree, plain code walks it, AI writes each topic’s lessons, plain code publishes them
  • AI in only two of the four stages; the mechanical stages always behave the same
  • Every topic is generated and checked against a strict format: the contract with the app
  • Written guidance for the model by age band, from a framework of cognitive levels
  • Small, independent topics: a weak one can be rewritten without touching the rest
2–4¢

The right model, not the cheapest

The challenge

  • Free forever means every cost matters
  • Children deserve lessons that are right, clear and pitched at their level

What we built

  • The cheapest model was tested first, and its output measured against that bar
  • Generation moved to a larger model in April 2026, once the quality difference was clear
  • At that model’s list price, a lesson costs an estimated 2–4 US cents to write
  • Writing all 2,786 lessons at that quality costs roughly $60–130: the quality is worth far more than the saving
74 / 97

Built with an AI pair programmer

The challenge

  • One engineer, many systems: a student app, a teacher records system, a content engine and three websites
  • Plain HTML and JavaScript with no frameworks, by design

What we built

  • One short rules file every AI session reads first: discuss before acting, change code only on explicit instructions
  • Hard constraints and non-negotiables in writing, so the assistant never guesses
  • Each folder documents itself; a status file records what is ready, blocked and decided
  • 74 of 97 code changes co-written with the AI, each one accepted by an engineer
Private

Safe for teachers’ data

The challenge

  • The teacher records system is multi-user: grades, attendance and students belong to the teacher who created them
  • No leaks between accounts, ever

What we built

  • Every read scoped to the signed-in teacher; every write checks ownership first
  • An AI code review and an AI security review before launch; the sign-in issues found were fixed and verified
  • A 53-check test battery against the real system: 28 cross-account attempts rejected, 20 simultaneous writes
  • Launched on 5 September 2026
0

Built for the oldest phone

The challenge

  • Children with poor connections, old devices and no tutor at hand
  • Families who should not pay with their data or their attention

What we built

  • Works offline: download a topic, keep playing, sync later
  • No passwords: an email address and a six-digit code
  • No ads, no profiling, no data sold
  • Light by design: no frameworks, small files

The timeline

From first commit
to public alpha.

Nov 2025 – Apr 2026 · App and content engine

A game first, then the content.

AreaWhat happenedThe number
AppStudent app built and stabilised; first beta on 2 December 2025Beta 1
EngineContent pipeline rebuilt as four clean stagesJanuary 2026
ContentTopics generated across three age levels, January to April733 topics
QualityGeneration moved from the smallest model to a larger one after reviewing output9 April 2026

A full catalogue of lessons and practice, aligned to the national curriculum, written in four months.

Sep 2026 · Teacher records system

From prototype to production in days.

AreaWhat happenedThe number
DataMoved from files to a real database, multi-user with per-teacher ownership1 owner per row
ReviewAI code review and AI security review of the new systemFixed before launch
TestingFull test battery against the real system53 / 53
LaunchAttendance, grades by period, reports for families5 September 2026

Teachers get a records system that does not get in the way, free, with their data kept theirs.

Now · Public alpha

Growing every week, in the open.

  • App and teacher records system live, content growing week by week
  • Teachers and specialists invited to review topics and flag the error nobody saw
  • The non-profit association is being constituted in Costa Rica
  • The “Un Millón por la Educación” campaign funds servers, content generation and reaching communities without connectivity

By the numbers

Small team,
whole curriculum.

Curriculum

172 of 173

Topic trees planned against Costa Rica’s official programmes

The rest follow SuperEscuela’s own learning path

Practice

54,820

Answer options written for the practice questions

Every question allows unlimited attempts

Teachers

28 / 28

Every attempt to reach another teacher’s records was rejected: each teacher’s data stays their own

Plus 20 simultaneous writes, none lost

Reach

Offline

Works on the oldest phone, without a connection, with no passwords

Progress syncs when the signal comes back

What it shows

  1. Let AI write, let code do the rest

    AI where judgment and language are needed; plain, predictable code everywhere else.

  2. Choose the model by measuring quality

    The cheapest model was the first tried and the first replaced. A few cents per lesson buys the quality children deserve.

  3. Write the rules down

    For the AI and for people: constraints, non-negotiables and status, read at the start of every session.

  4. Review with AI, decide as people

    AI reviews the code and the security; engineers decide what ships.

  5. Stay open to scrutiny

    Method, content and decisions can be reviewed and questioned by anyone.

Have a mission
that needs to scale?

Non-profit or business: tell us what you need to produce or run at a scale your team cannot reach by hand. We will show you where AI fits, and where it should not.

Prefer to write directly? enable JavaScript to see the address

Talk to an engineer

No sales theater. Tell us where your operation feels slow, repetitive, or difficult — an engineer reads every message.

Your message goes straight to our engineers at our address.