Diktafone
Talk for two minutes about someone you just met. The phone transcribes it, extracts a structured profile, and never sends a byte anywhere for inference.
2.4 GB
Of model weights running on the phone
0
Fields the AI can write without my approval
105
Jest cases, run on every push
Why I built it
I meet a lot of people through college, freelance work and student organisations, and I remember faces. What I lose is the detail. Clay.earth solves that by ingesting your email and calendar into someone else's cloud, which defeats the point when the notes are candid ones about friends.
One loop, two minutes
Tap record, talk, walk away. Whisper transcribes on the handset and Qwen3-1.7B turns the transcript into structured fields. Capture writes to disk before it touches the network, so it works with no signal at all.
The AI proposes, I dispose
The model is never allowed to write to the database. Every field arrives as a draft carrying its confidence and the sentence of transcript it came from, and I approve, edit or reject one field at a time.
One queue for everything
Pending drafts, tier-change nudges, stale-relationship prompts, duplicate suggestions and unresolved conflicts all land in one inbox. Dismissing a nudge snoozes it and never deletes what it pointed at.
Everything runs on the phone
2.4 GB of weights, CPU only, one model context resident at a time. The pipeline is a serialized chain, so Whisper is freed before Qwen loads. That is what stops an 8 GB handset from OOM-killing the app mid-run.
More screens
5 captures
