Inspiration

A comprehensive geriatric assessment asks people to describe many parts of their lives. We wanted that experience to feel more like a conversation than a long form, while keeping the information useful for the clinician who reviews it. Elderwise pairs a voice-guided assessment with an expressive virtual Reachy Mini companion.

What it does

Elderwise guides a conversation across 14 comprehensive geriatric assessment domains. A virtual Reachy Mini avatar gives the voice interaction a visible companion. The assessment keeps transcript provenance and supports corrections, then produces a draft report for clinician review.

The report is a review aid, not a diagnosis. This submission demonstrates a browser-based virtual robot, not a deployed physical robot or a clinically validated medical device.

How we built it

The application uses nvidia/Nemotron-3.5-Super-VL through the Nebius Token Factory chat-completions API for assessment chat, parsing, voice-session summaries, and clerking. The integration is prompt-configured. ElevenLabs provides the separate voice-agent runtime. The interface pairs the conversation with a virtual Reachy Mini avatar and retains transcript provenance so assessment content can be reviewed and corrected before it becomes a draft report.

An optional Tavily integration can search public information.

Challenges

The central design challenge is preserving the distinction between what a person said and what an AI inferred. Transcript provenance, correction support, and clinician review make that distinction visible. Another challenge is making a structured assessment conversational without presenting a prototype's output as a diagnosis.

Accomplishments

The working demo brings together an expressive virtual companion, voice interaction, a 14-domain assessment, transcript provenance and corrections, and draft reports for clinician review. NVIDIA Nemotron runs through Nebius Token Factory as part of the assessment experience.

Design takeaway

A useful assessment assistant needs more than fluent conversation. The reviewer must be able to understand where the information came from, correct it, and distinguish an AI-generated draft from a clinical conclusion.

What's next

The next evaluation questions are conversation quality, accessibility, language support, and the usefulness of the draft report to clinicians. A physical Reachy Mini implementation would be future work. No physical deployment or clinical validation is claimed for this demonstration.

Source availability

The linked public repository is a limited demonstration companion, not the complete Elderwise production source. Additional repository and implementation details are available to the judges on request.

Public companion: https://github.com/yamyr/elderwise-nebius-demo