SoulSync explores multimodal AI that monitors tone, pacing, and contextual cues to support elders and caregivers. We built lightweight models that run fully on-device to preserve privacy, and tuned them for low-light and low-bandwidth environments where typical cloud assistants fail.
We added adaptive prompts for different cognitive states, offered optional journaling summaries that can be shared with clinicians, and designed escalation paths that notify family when risk signals like sentiment drift, agitation, or prolonged silence appear. The system emphasizes consent, transparency, and clinician-in-the-loop review before deployment, so that automation augments human care rather than replacing it.
A key learning: usability beats model novelty. Care teams preferred predictable, explainable actions, like asking a grounding question every few minutes, over complex opaque behaviors. This shaped our evaluation rubric for safety, comfort, and trust.
Related evidence