Papuguinho
AI & ML interests
None defined yet.
Recent Activity
Papuguinho
Papuguinho is a free, open-source Augmentative and Alternative Communication (AAC) system developed at IFSP Campus Jacareí (Instituto Federal de São Paulo).
It helps non-verbal individuals — children with autism, cerebral palsy, Down syndrome, and other conditions — communicate through pictogram-based boards, entirely in Brazilian Portuguese.
Products
🦜 Prancha Digital (Papuguinho App)
A cross-platform AAC app (Android, iOS, Web) for building and using pictogram communication boards. → www.papuguinho.com
🖨️ Estúdio Papuguinho
An AI-powered web tool that generates printable pictogram boards from natural language prompts. Powered by the models in this org, it searches ~14,000 ARASAAC pictograms semantically to assemble a board in seconds. → estudio.papuguinho.com
Models
| Model | What it does | Size |
|---|---|---|
| aac-board-generator-moe-636m-ptbr | Generates pictogram labels and structured board data from natural language — Mixture-of-Experts LM | 636M total / ~212M active |
| aac-pictogram-retriever-ptbr-GGUF | Semantic search + reranking over the ARASAAC catalog in pt-BR — GGUF for llama.cpp | 157M |
Both models run on CPU, without a GPU. The full pipeline operates with ~2 GB RAM on any modern machine or low-cost VPS.
Why it matters
Brazil has an estimated 45 million people with some degree of disability. For those with Complex Communication Needs (CCN), commercial AAC solutions are often prohibitively expensive or unavailable in Brazilian Portuguese. Papuguinho provides a high-quality, free alternative built specifically for this community.
The AI pipeline achieves performance competitive with large commercial cloud APIs (OpenAI, Voyage, Gemini) while running entirely locally — ensuring privacy by design and zero ongoing cost for institutions.
Open source
All models, datasets, and tools are open source:
- 📊 aac-pictogram-retrieval-ptbr — Evaluation benchmark
- 🏋️ aac-pictogram-retrieval-ptbr-training — Training data
- 🌱 colibri-embed-ptbr — Base pt-BR embedding model
- 🔧 gemma3moe-builder — Tool for converting dense Gemma-3 models to MoE
- ✂️ embedding-vocab-trimmer — Vocabulary pruning for multilingual embedding models
License
Models are released under the Gemma license. ARASAAC pictograms are authored by the Government of Aragón under CC BY-NC-SA.
Developed at IFSP Campus Jacareí · www.papuguinho.com