Jnana Kosha — Digital Library

The hardest part wasn't the app — it was moving years of content out of three systems without losing a byte or inventing a link that was never there.

We built a five-stage, offline-first pipeline: export, combine, prepare, import and an independent reconcile. Preparation touches no database or network and emits a hashed plan. Import dry-runs by default, writes in a single transaction, is idempotent and refuses to run against production by accident. Reconciliation then re-checks counts, hashes, identities, chapter order and every media URL — read-only.

Audio albums are never matched to books by title. Every association is an explicit, reviewed decision.

Recitations, generated in-house

Many stotras had text but no recording. We generate recitation audio with our own Sanskrit and Kannada voice pipeline, and treat every clip as editorial content, not a byproduct: each recording is tied to the exact text version it was made from by hash, carries its timing data and version, and stays unpublished until a reviewer approves it. AI provenance is stored with every clip and disclosed in the app.

Details that took real care

• Language modelled separately from script — a Sanskrit hymn in Kannada letters is not a Kannada translation
• Text preserved byte-for-byte and verified by hash
• Media files verified by signature with range requests, not trusted by URL
• Persistent mini-player across tabs, speed control and saved progress
• In-app corrections queue with idempotent submissions
• Consent-gated analytics — off by default, de-duplicated and pruned after 90 days
• Tested at 200% text size, with reduced-motion and dark-mode contrast checks