← CodeswitchReading level: normal humanunnecessarily precise
Voicebook · Architecture

Family memories that never leave the house.Zero-egress family archive. Local inference only.

The recordings of a family's everyday life are the most private data there is. So the whole system runs on a computer at home: no cloud upload, no third party in the loop.Capture on iPhone, transcribe and index on a Mac in the house, retrieve with RAG over local files. No audio, text or embeddings cross the trust boundary.

Runs at homeOn-prem (a closet)No cloud uploadEgress: 0 BAsk it questionsGrounded retrieval
transport: private businference: localremote: WireGuard meshtrust: you
9:41 ●●● 100% New Recording Saved to this device only 00:42 Dinner table · Sunday
Step 1Recorded on the family's own phoneiPhone → local capture queue
How it fits togetherReference architecture

One home. One private bus. Nothing outside.One boundary. One bus. Zero egress.

TRUST BOUNDARY · YOUR HOMETRUST BOUNDARY · ZERO EGRESS BY DESIGN iPhone Voice memosOn-device capture 1 Mac Local AI + recallLocal LLM + RAG 3 Screens Optional replayRead-only clients · THE BUS one private lane every device and app hands work through job tickets in · results out · idempotent · no inbound ports 2 External drive Entries, moments and clipsPlain files, local storage 4 Nothing here is uploaded to a cloud service. Not for storage. Not for AI. Private tunnel encrypted, device to deviceWireGuard mesh, mutual auth 5 Away from home Your signed-in phoneauthenticated client device The public internet never seesthe archive or the model.
2The bus

Every part talks over one private lane.Components never call each other. They file tickets.

The phone, the Mac and the drive never call each other directly. They pass work across a single bus that lives inside the house, so each piece stays small, swappable and easy to audit.Producers enqueue job tickets; consumers claim, process and write results back. Loose coupling means any stage (transcribe, clean up, index) can be swapped without touching the rest.

  • Work goes in as a ticket, comes back as a resultAt-least-once delivery, idempotent handlers
  • Any piece can be replaced without touching the restSwap a model, keep the contract
  • The bus has no door to the outside worldNo inbound ports. Nothing listens on the internet.
A glowing home with data swirling safely inside
Trust boundaryEverything inside stays inside
3Local AI

A model that listens in the house, not in the cloud.Inference at the edge. The edge is your living room.

A language model running on a Mac at home turns messy, multilingual family chatter into clean, searchable entries. It picks out the words the kids are learning, the stories worth keeping, and who said what.Local speech-to-text and an open-weights LLM segment audio into moments, tag speakers and languages, and write embeddings to a local index. Code-switching handled in the same pass.

  • Mixed-language speech understood in placePer-segment language ID, bilingual transcripts
  • New words and moments tagged automaticallyMoment mining + vocabulary extraction
  • No audio or text leaves the machineAirgap-adjacent, minus the drama
Tangled multilingual speech bubbles resolving into a clean dictionary
Before → AfterChatter into a dictionary
See it workThe review console

Parents stay in charge of what gets kept.Human-in-the-loop on every write.

Review screen showing a recorded moment
1 · Review Listen, then keep, skip or save for later.Triage queue with accept / log-only / reject.
Polish screen showing a cleaned-up transcript
2 · Polish Clean up the words, in two languages.Per-speaker cleanup with translation.
Search screen finding a moment by a word
3 · Find Type a word, get the moment back.Keyword + semantic retrieval over the archive.

Illustrative screens with invented entries.

4Archive + private recall

Ask your own archive anything.RAG over files you can open in Finder.

Entries sit in a plain folder on a drive at home. When someone asks a question, the model retrieves from those files only, so every answer points back to something the family actually said.The corpus is plain files on local storage. Queries retrieve from that corpus only and cite the source moment, so answers are grounded rather than generated from thin air.

  • Answers grounded in the family's own filesCitations to source moments
  • Plain files you can open, back up and moveNo proprietary store, no lock-in
  • Backed up on the family's own termsYour 3-2-1, your rules
2026-03 · First bike ride Dinner table · 0:42 Family Journal entries · moments · language log on your drive plain files
IllustrativeInvented example entries
5Private tunnel

Reach home from anywhere, with no open door.Remote access without a single open port.

Away from the house, a phone connects to the home computer through an encrypted device-to-device tunnel. Only the family's own signed-in devices can see it, and it never shows up on the public internet.A WireGuard-based mesh (Tailscale) authenticates each device and builds direct encrypted paths to the Mac. No port forwarding, no public address, no cloud relay of your content.

  • Replay and search the archive on the goSame bus, different network
  • No public address to attackAttack surface: your front door, not your router
  • A lost device can be removed from the networkDevice revocation from the admin console
A phone being used on a bus
Isometric home with a private connection to the outside
EncryptedDevice to device
The safest place for a family's memories is the one nobody else has the key to.Threat model, summarized: if it never leaves, it can't leak.