Inspiration
In Argentina, a student with a disability in a mainstream school has an individual inclusion plan (PPI, Resolution CFE 311/16). In a UNICEF Argentina study of 20 schools in 5 provinces (2020), only 53% of the education actors consulted said they had clear guidelines for monitoring those plans, and only 34% of the students said they had been invited to any stage of their own plan. I live in Aguaray, in Salta, northern Argentina. Hermano works on that gap: the day-to-day follow-up, and the child's own voice.
What it does
The child picks a companion, hermano (brother) or hermana (sister), that walks them through the school day on the phone: a first/then cue for every block, a heads-up five minutes before each change, calm-down options and an "I feel nervous" button on every step, and three faces to say how each step went.
Every adult around the child gets their own view, only about their own children: the teacher gets practical notes, one reinforcement idea per class and the new ideas to approve; the family gets plain language and the same ideas to approve; the specialist gets a structured log; the principal gets counts only.
Why it only works with Qloo
Without Qloo, a child would get the two or three favorites an adult typed, over and over. That is exactly what Hermano falls back to when Qloo does not answer. With Qloo, Hermano looks up each child's interests with /search and asks /v2/insights for related books, artists, TV shows, video games and movies, with filter.content_rating for the child's age. Ideas become rewards after classes, movement breaks, calm music and reinforcement ideas for teachers. Rejected ideas are excluded from the next Qloo request using filter.exclude.entities.
A real example from the live demo (12 October 2026): Leo, an invented 8-year-old, likes Bluey and Minecraft. From Qloo's answers, the agent suggested Dug Days to his family because it "fits the child's Minecraft taste", and Onward because it "pairs well with the child's love of Bluey".
The agent and safety
A language model (Qwen3.8-27B on Groq) plans each child's day through get_day, get_motivators and submit_plan. It selects adult-approved ideas and suggests up to two new Qloo ideas with a reason naming the child's taste. Code checks choices: no reward after recess or dismissal, nothing adults did not approve, no Qloo idea twice in a day, no health words or contact data, and no reason that names none of the child's tastes. The model sees a placeholder, age, blocks and titles, not the child's name or supports. If unavailable, the system falls back to gpt-oss-20b and then a rule-based planner.
Nothing new reaches the child without an adult's OK. An age gate runs before an adult sees an idea. Movies and shows need a suitable rating; games use the stricter of Qloo's ESRB rating and the team's table, and under 13 an unrated game is not proposed. Under 13, a new artist or book needs a Qloo tag for children or families. Real runs exposed an age-gate problem with a T-rated game and an unsuitable artist; those holes were fixed.
Privacy by design
No cameras or biometrics. Supports and exercises come from closed catalogs, without diagnosis or free-text notes. Only interest names go to Qloo; names, health words and contact details are refused. The demo shows every request sent to Qloo that day.
Try it
Live demo: https://hermano.onrender.com. The school, children and adults are invented: Leo (8), Sofi (11) and Mati (14). The free host may take about a minute to wake. The agent plans in the background.
- Open Leo under "The child's view" and tap "I feel nervous" on a step.
- Open Leo's family (demo), see new Qloo ideas and tap Yes on one.
- Return to Leo's view to see the approved idea. Each visitor gets their own demo; "Start my demo again" resets it.
- Open the 3rd grade teacher, specialist and principal pages.
- Open "What was sent to Qloo" and switch to Spanish at top right.
How we built it and challenges
Python with the standard library, English and Spanish, hosted on Render. The prepared build had 189 automated tests. An independent AI audit produced four findings that were fixed with tests and checked on the live demo. Qloo's video_game type returned 400, so we use urn:entity:videogame; real Qloo responses exposed age-gate holes mock data missed. Groq's edge returned 403 for Python's default user agent. An audit found that an adult's Yes did not reach the child's web view; it was fixed and tested.
Idea and product decisions: Jorge Abraham Mendieta. Code written with AI assistance.
Known limits and next steps
Everything uses invented data; Hermano has not been tested with real children or schools. The rating filter excludes titles without a US content rating, including some local Argentine shows; under 13, books and artists without Qloo tags are also excluded. In the web demo each day is treated as a first day, and after an adult's Yes the child is replanned by rules rather than by the model. Next: a pilot with a school team in Salta, families and specialists; local content ratings; and reward rotation across days in the web app.
Log in or sign up for Devpost to join the conversation.