File 06 · Nebius × NVIDIA Global AI Hackathon · Personal AI · Oct 2026
Vox
Talk to your day. You ramble, correct yourself halfway through, and Vox keeps only what counts, with no forms, date pickers or dropdowns.
- ~0.5 severyday commands on Nemotron 3 NanoMeasured
- 3model sizes, routed per requestMeasured
- 37tests: safety, undo, dates, routing, securityMeasured
You said four things and cancelled one mid-sentence. Vox kept the three that count and gave each a real date. There is no dentist task, because you said “no wait, cancel that.”
Each request goes to the right-sized model: Nano for everyday commands, Super for long brain-dumps and meeting summaries, Ultra for planning. An activity log shows which one answered and how fast.
It also drives you there: map, route, spoken turn-by-turn, and a hands-free co-pilot mode with a road mix that quietens whenever Vox speaks.
What it refuses to doLetting the model touch your data directly. It answers with a JSON plan, and only validated actions reach your tasks.
How it runs
- ListenEnglish, French or Arabic speech, messy and out of order.
- RouteNano, Super or Ultra, depending on what you asked.
- PlanThe model returns a JSON plan: add, update, complete, remember, navigate.
- ValidateNothing changes until the plan passes validation.
- Follow up“Did you do it?” Answer by voice.
Honest caveatLatency figures are the project’s own typical values, shown per model in the app.
Next fileDeflake →