VEXO KIN

Vexo Kin — your band remembers who you love. It listens, builds private friend profiles, and nudges you at the right moment. The emotional layer Vexo was missing.

Comment

Vexo Kin — The Emotional Memory Layer

What Inspired Us

Every AI wearable today is built to make you a faster worker. None of them help you become a better friend.

We noticed Vexo already knows what you need to do — but it has no memory of who you love. Your friend mentions her interview on Thursday. Your brother drops that he's been stressed. It disappears into thin air.

We asked: what if the band just remembered?

What We Built

Vexo Kin is a passive emotional memory layer on top of Vexo. It listens to everyday conversation, extracts implicit signals — wishes, upcoming dates, stress, promises — and builds private friend profiles only you can see.

When Thursday comes, it nudges you: "Maya's interview is today."

No typing. No app. Just showing up for people.

How We Built It

  • Band mic → Deepgram nova-3 for real-time speech transcription
  • OpenAI GPT-4o extracts structured signals from raw transcript
  • Band key-value storage (242KB persistent) stores friend profiles
  • Python backend ties the pipeline together
  • Bubble UI — you in the center, friends orbiting as nodes

The extraction schema:

{
  "person_mentioned": "Rachel",
  "signal_type": "date",
  "content": "job interview Thursday",
  "sensitive": false
}

Sensitive topics — health, money, religion, politics — are automatically discarded. Privacy is not a feature, it's the default.

Challenges We Faced

Python environment on macOS — pip 26 has a macOS version detection bug (ValueError: invalid literal for int()). We lost an hour to this before bypassing ensurepip with get-pip.py.

No speaker identification on the band — the mic captures everything but can't tell who's talking. We solved it with a simple UI toggle: "talking to: [name]" — honest for a demo, and the right UX for a privacy-first product anyway.

242 bytes per storage key — friend profiles had to be aggressively compressed to fit. We store only signal summaries, not raw transcripts.

4 hours. That was our entire build window.

What We Learned

The gap in AI wearables isn't productivity — it's intimacy. The hardest engineering problem wasn't the stack. It was deciding what not to store.

Built With

Share this project:

Updates