
Vox AI
- Python
- Supabase
- pgvector
- Gemini API
- GitHub Actions
- RAG
[Problem]
A support and safe-information assistant for the LGBTQIA+ community in Salvador. The initial MVP using spreadsheets, repo JSON, and local sentence-transformers downloaded the model on every Streamlit boot and did not scale.
[Architecture]
Decoupled Python backend on Supabase (PostgreSQL + pgvector) with RAG, Google Gemini API, and Anonymity by Design (zero PII). GitHub Actions CI/CD running Gatekeep and Semgrep.
[Outcome]
~80% lower RAM usage by migrating from local inference (sentence-transformers) to pgvector + REST APIs, with higher concurrency and ongoing open-source maintenance in production.
Python · Nov 2026
- RAM Usage
- -80% (pgvector + REST)
- Vector Store
- Supabase · pgvector
- CI/CD Security
- Gatekeep + Semgrep
- License
- Open Source (GPL-3.0)

