Yaad — Every Story, in Their Own Voice

Watch

The 52-second tour

Demo with a fictional narrator, voiced synthetically. The transcripts, interviewer replies, analysis and booklet are the app’s real output.

The Problem

A lifetime of stories, never written down

Our parents and grandparents carry stories no one else knows — where they grew up, how the family came to be, what life was like before. Almost none of it is ever recorded. Writing a memoir is too much to ask, typing is a barrier for many elders, and when they are gone the stories go with them.

What I Built

A voice note, a good listener, and an archive that lasts

Yaad works the way elders already talk to family: press and hold, speak, release — just like a WhatsApp voice note. Each recording is transcribed, and an AI interviewer replies the way a thoughtful grandchild would: it reflects on what was said, asks a single gentle follow-up, and speaks the reply aloud in a voice the family chooses. For the family, an analysis maps the people, themes and time periods mentioned and suggests what to ask next, and a booklet compiles every story in the narrator’s own words.

Technical Highlights

Built for the next thirty years, not the next release

Capture never depends on the network

A recording is saved in the browser the instant it stops, then uploaded by a background sync that retries — a dropped connection can’t lose a story.

Raw audio is write-once

Original audio and its verbatim transcript are written once, checksummed with SHA-256 and made read-only; everything else is labelled as derived and regenerable.

Files are the record

The archive is plain files on disk; SQLite is only an index, and one command rebuilds it from the files at any time.

An interviewer with restraint

One question at most, reflect before asking, an occasional pure reflection with no question at all, and always in the narrator’s own language.

Never rewritten

The booklet is assembled from verbatim transcripts by a template, not by AI — so it can never be mistaken for a rewrite of the narrator’s words.

The family’s own keys

Each family brings its own AI provider keys, encrypted at rest with AES-256-GCM and used only for their own recordings; no default key is built in.

Stack

Under the hood

Frontend	React + Vite, installable mobile-first PWA with press-and-hold recording
Backend	Node + Express, flat-file archive with a SQLite index, background job workers
AI providers	Whisper transcription, Claude or GPT interviewer, OpenAI text-to-speech — behind one provider interface
Hosting	Self-hosted with Docker and Caddy on a persistent disk — deliberately not serverless

Status

Where it stands

Fully built through all seven planned stages, from capture to the family view. Yaad1 carries the same idea into a multi-family service on shared hosting, adding guided oral-history projects.