Sansara

The software runs nine distinct AI agents on a mobile device, each with a specialized role such as empathy, insight, or growth guidance. These agents interact through a Blackboard consensus loop, proposing, critiquing, and refining responses before delivering them to the user. All inference is performed locally via ExecuTorch, eliminating the need for cloud servers and keeping data on the device.

Designed for personal wellness and reflective journaling, the system processes multimodal inputs like voice stress and facial tension to inform its replies. The consensus architecture aims to avoid single‑model hallucinations and provide a more nuanced, non‑prescriptive presence that mirrors the user’s own thoughts rather than offering directives.

The approach is experimental and targets users who want on‑device, privacy‑preserving AI assistance for emotional self‑exploration, with a focus on continuous learning that adapts the model to the individual over time.