Vasyl Stanislavchuk

Jadwigo is a partner for parents in their child's learning. Today it covers the first 4 school grades, in Polish and English. The account is the parent's. A parent sends a photo of the homework, a few notes on what their child struggles with, or a message from the teacher. Jadwigo builds a lesson for that child in about two minutes. The child plays it on screen, or the family prints it and works on paper together. Problem: Parents want to be part of their kids' learning. Time, energy and each child's own needs get in the way. Time is short. Homework help happens in the gaps of a busy day, often while dinner cooks. Helping takes energy. We often have to relearn a topic before we can explain it. Every kid is different. The whole class often gets the same homework, and each child makes their own mistakes. Each one needs practice on their own gaps. We live this at home. I'm a dad of 2 boys, and we recently moved to Poland: new language, no grandparents, few friends. Khan Academy sets the pace, but its questions aren't about my son's challenges. Duolingo runs on points and screens. Kids love Kahoot in class, but it doesn't follow them home. Solution: AI changes what a parent can do in one evening. It can read a photo of homework, a parent's notes or a teacher's message, and see where the child struggles. Then it writes practice for that one child. Jadwigo uses Meta's Muse models for this. The model doesn't write a lesson from scratch. It picks questions from a bank of Khan Academy's maths exercises with RAG, using hybrid search and a reranker. The questions stay as Khan published them, so the answers are Khan's, not the model's. The model lays them out in Jadwigo's 29 question types, which are already built as interactive screens. Then Jadwigo checks the lesson for correctness and quality. Each parent message starts an agent run, a chain of steps. Each step is one Muse call plus the tools it asks for: search the bank, draft, edit, check, publish. A worker on AWS Lambda runs the steps, and the parent watches each one happen. Checking is code, not a second opinion from the model. It validates every question against its type's schema and refuses harsh words, such as a bare "wrong". Each child has a learner knowledge base. It maps skills to the school curriculum and tracks mastery per skill from every judged answer. From that history, code works out a brief of what to reteach, review, start or recheck, and the next lesson is built from it. On request, Jadwigo runs a "dream". Code computes the child's progress, and the model writes it up for the parent, citing the answers behind it. The parent chooses what's next. The same tools are served over MCP, so parents can use Jadwigo from ChatGPT, Claude or Muse. Today I built Jadwigo for my son David. His class plays it now, and their teacher has a cameo in the lessons. Parents from the class write to me that they love it. His school is next. Maybe after, it's you? https://jadwigo.pl/en/