Approach
Systems that endure are built step by step.
OASI systems share around 70% of their infrastructure through a common library — @oasi/core-ai. Every new vertical inherits what has already been solved: streaming voice, RAG, telemetry, hardening.
I
@oasi/core-ai
The technical backbone: local LLM orchestration, transcription pipelines, vector indexes, security models, observability. We write it once, every system uses it — Atena and Avatar.
II
On-premise, where data doesn't move
When an enterprise customer cannot — or will not — let their data leave, infrastructure must move closer to the data. Physical servers, local models, a defined security perimeter. It is the necessary condition for serious AI on confidential content.
III
Cloud, where it makes sense
For what is public, scalable, distributed, cloud is the right tool. But the choice is never default: every component lives where the nature of the data demands. Never a compromise on confidentiality for deployment convenience.
Open source
Two forks we maintain, adapted for Italian.
Voice components built on others' work. We maintain them, optimize them for Italian use, and keep them public. MIT-licensed.