Mentora

In my own job the shift already happened. I do not type most of the code anymore. I direct agents, review what they write, and keep the calls that matter. Lawyers went from research by hand to first drafts and case review with AI at their side. Radiologists read scans with a model flagging what to look at. Designers, accountants, marketers, support teams, anyone who writes for a living: the work moved from producing to directing, judging and verifying.

Education did not move. The students did. They already draft, translate and study with AI every day. The classroom around them still runs on the same exams, the same weekend lost to grading, and the same single percentage that says nothing about what a student actually misunderstood.

What schools should do

Not ban it, and not surrender to it. The job of a school is to prepare students for the world they are going to work in, and that world runs on directing and checking machines, not on competing with them. That means two things at once. Teachers get their time back from the repetitive part, and use it for the part only a person can do. And assessment stops being a number and starts being evidence: what does this student understand, topic by topic, and what should we do about it tomorrow.

With care

UNICEF’s Digital Education Strategy for 2025 to 2030 asks for technology in schools that is human-centred, equity-driven and evidence-based, with teachers at the centre and not on the sidelines. UNESCO’s guidance on generative AI in education says the same in different words: a human-centred approach, teachers trained and in charge, and real protection for students. I take that as the blueprint, not as a footnote.

In Mentora it translates into rules I do not bend:

• The teacher stays the teacher. AI drafts. The teacher edits, replaces and approves. Nothing reaches a student that a teacher has not reviewed.
• Grading shows its reasoning. Written answers are graded against the teacher’s own model answer, with a written explanation, and anything uncertain falls back to manual review.
• Students are always free, and feedback speaks the student’s language, not the teacher’s.
• Evidence over vibes. Every result becomes per-topic evidence, so the conversation is about what to reinforce, not about who got a 6.

Mentora is an independent product, not affiliated with or endorsed by UNICEF or UNESCO.

What Mentora does

A teacher describes what they taught: subject, topics, how hard it should be, how many questions, multiple choice or written or a mix, and optionally the source material itself, pasted in. The exam drafts in front of them, question by question, with options, correct answers, explanations and a rubric. They edit whatever they want and publish it to a virtual classroom, with an availability window, a time limit and randomised questions if they like. Students join by email invitation or a six-character code.

While students sit the exam, the teacher sees it live: who opened the paper, progress filling in, cards flipping on hand-in. Multiple choice grades instantly. Written answers are graded by an LLM against the teacher’s model answer, with reasoning attached. Then every result folds into a picture of each student, topic by topic, weighted toward what happened recently.

On top of that sit the reports: a debrief after each exam that the teacher can act on tomorrow, a classroom pulse that reads like the note a good head of department would leave, and for school directors a morning briefing plus a live board of the whole school. The signals underneath are deterministic: teaching gaps, topics that are decaying, students at risk, momentum. The AI writes the commentary, it does not invent the facts.

It is built for independent teachers first, then for schools that want the same across every classroom. Spanish and English, with Latin America and Argentina as the starting point.

How it is built

A React and TypeScript single-page app talking to a Node and Express API, with Drizzle on PostgreSQL underneath. Sign-in is Google only, with server-side sessions. Realtime is one WebSocket per tab. The LLM work sits behind a thin service layer, so the model is a configuration value and usage is metered per account. Hosting is Vercel, email is Resend, and a daily job produces the reports.

It is live at mentora.ar and I am still building it. Billing and the school tier are the current work. If you teach and want to try it, write to me.