Gabriel Costa

Case study

A maths app for secondary school students. Students photograph their working on an exercise and the AI grades it, explains the mistake and prepares them for the national exam.

Type
Own product
Role
Founder: product, design and engineering
Platforms
iOS, Android
Started
2025
Status
In development
Wolfi home page
26

server functions

4

official sources monitored

5

stages in the agent workflow

Context

Wolfi started as my undergraduate project at the University of Beira Interior and is becoming a product.

What it does

01

Grades the working

Students photograph a solved exercise and get grading and feedback on the reasoning, not just the result.

02

A tutor on call

Answers questions about the exercise the student is on, at any hour.

03

Exam practice

Exercises, quizzes and past national exams, organised by the official curriculum.

04

A plan to exam day

A study plan that follows the student, day by day, up to the exam date.

The curriculum and exams change. Instead of reviewing content by hand, I built agents that detect changes and propose what to do. Nothing reaches students without approval.

  1. 01

    Monitoring

    An agent follows the official sources and records what changed.

    IAVE, DGE and DGES

  2. 02

    Impact

    Each change is mapped to the curriculum, and content gaps are identified.

    Curriculum analysis

  3. 03

    Proposals

    Changes become proposals, each with its own risk level.

    Risk classification

  4. 04

    Review

    I approve, reject or request changes. Every decision is logged.

    Human review

  5. 05

    Execution

    Only approved, low-risk proposals go ahead. The rest stays blocked.

    Controlled execution

Technical decisions

  1. 01

    Agents on a short leash

    Agents propose, but only approved, low-risk work proceeds on its own. Everything else waits for a person.

  2. 02

    AI runs on the server

    Keys never ship in the app, and models and instructions can improve without releasing a new version.

  3. 03

    AI costs under control

    Each AI feature has per-student limits, and a repeated request is never counted twice.

Status

Stack

App
Expo
React Native
TypeScript
NativeWind
Backend
Supabase
Postgres
Edge Functions
AI
Google Gemini
OpenAI
Operations
Next.js (Wolfi Ops)
RevenueCat