Eligiendo Mi Camino
A World Bank and DRELM program in public schools in Lima, winner of the World Bank Group's Innovation Award. My work: turning the specifications and the initial concept into two AI tools running in real classrooms.
Ten out of a hundred reach the level. Three out of ten keep studying.
In Peru's public system, only 10% of students reach the expected level in maths, and barely 3 out of 10 make it to higher education. The program tackles both ends at once: maths reinforcement with an AI tutor, and career guidance to decide the path after school. The goal: that every student in their final year of secondary school goes on to study at a university or a technical institute, or moves into work.
A sea of specifications and a prototype of happy paths.
The program arrived co-created: the World Bank brought the concept, the pedagogical rules, the data and a concept prototype. All of it valuable, and all of it incomplete for a real classroom. The flows covered the happy path: no edge cases, no error states, none of the logistics of an actual school session.
My work was that gap: breaking the specifications apart, restructuring the flows and solving what was missing. What happens when a student joins halfway through the program, when a teacher needs to step in, how an interrupted session is picked back up. And documenting all of it for the handoff to development.
And that gap had to be crossed twice, because the program is two tools that live together across the whole school year. Domina Matemática is the tutor that reinforces what is covered in class; Orientación Vocacional, the coach that accompanies the decision about what comes after school. They share students, teachers and the front door, but their session logic is the opposite. One is practised when it comes up and can be repeated; the other advances in steps that don't go back.
A tutor that asks before it answers.
The maths tutor works through Socratic tutoring: when the student gets stuck, Doc doesn't hand over the answer — it helps them find where the error was. The question bank is aligned to the national curriculum (four competencies, sixteen topics) and the diagnostic reads by competency and topic, not by score.
A coach that guides: the eight-step journey.
Career guidance isn't a loose test: it's an 8-step journey across the school year. Self-knowledge, interest test, myths and facts with cited sources, post-secondary paths, exploring 130 occupations across 12 sectors, research with AI, a family decision and a downloadable actionable plan.
That map is the spine of the program, but every step is a different piece. An interview with the coach, a 67-question test, an explorer with search and favorites, a plan you download. What ties them together is that each one leaves something behind (a profile, a list, a decision) that the next one uses as input.
Step 3 is the one that best shows the program's pedagogical bet. Instead of giving a talk about careers and salaries, it puts the student in front of statements they have probably already heard at home. And it answers with data and the source cited, right or wrong. It never punishes a wrong answer: closing the step isn't about scoring well, it's about leaving with data you didn't have before.
Three doors for three roles.
- The student signs in with their national ID (DNI), no emails, no account creation: the friction of a classic sign-up doesn't survive a final-year classroom.
- The teacher monitors their section live: who is connected, which step each student is on and who hasn't logged in for days.
- Program supervision sees overall progress by school, and an inbox centralizes the reports that used to circulate over WhatsApp.
On top of the system that already existed, not a new one.
Two different tools (one for maths, one for career guidance) had to feel like the same product from day one. The base was the uDocz design system: buttons, inputs, cards, tables, navigation, feedback. None of it was redesigned, and that's the decision: a program starting in 110 schools at once is not the place to debut a system.
On top of it we built what the program did need and the system didn't have: a layer of its own components. The step card with its four states (pending, current, done, locked), the eight icons of the journey, the sidebar with its progress, the coach's bubble, the plan table. All with variants, so a completed step renders the same way in the sidebar, in the home map and in the step's own card.
And there is a navigable prototype in code (HTML, CSS and JavaScript, with synthetic data) that carries part of that design. It doesn't replace Figma. It's there to walk the complete flows with the three doors, test the path of a real session and hand development something that can be opened and used, not just looked at.
Changing the product without spending the client's trust.
With an institutional partner every change costs trust, and that currency doesn't come back. So before touching anything, I classified: every proposed change went into a bucket according to its risk to the program. The bucket decided the treatment: whether it was done, proposed or raised with the partner.
The result: we could iterate fast where it was safe, and walk into every conversation with the partner with a short, clear list of the decisions that really were theirs.
First phase completed, and a public program you can read about.
The first phase ran from March to July 2026 across 110 public schools in Lima, with 6,600 students and around 400 teachers trained. The program is public: these sources aren't mine, and that's why they count: