Mentezinha
An AI for kids that knows when to say no.
01The problem
With children as the users there is no acceptable failure rate, and "the model usually behaves" is not a safety plan.
02What I built
I built it for my son's homeschooling, and now other families use it. Each child has a retrieval index on Google's managed vector search that grows beyond what the parent configured, accumulating understanding across conversations so the next one is better.
03The decision that mattered
The guardrail is a validation layer with the authority to refuse, not a line in a prompt. A refusal is treated as a correct outcome, not an error, because with children it often is the right answer.
04Receipts
- RAG per child on Google Vertex AI with managed vector search
- Personalization that accumulates across conversations, not generic document RAG
- Guardrail that can refuse, enforced outside the prompt