Public URL: https://video-ai-digest.vercel.app/
This static, browser-local companion to the author's Node/Python media app reads TXT, SRT and WebVTT. It preserves timed cues, searches literal phrases, selects exact excerpts, separates user notes, fingerprints original text with SHA-256, and exports Markdown, text and a re-importable JSON workspace. Optional localStorage saving is explicit and device-local. It does not transcribe audio, fetch media URLs, call an LLM, generate translations or upload transcript content. CSP disallows network connections from page scripts.
The previous static demo is retained under archive/, excluded from deployment. The old demo.js is no longer loaded. No public control claims backend functionality that is unavailable here.
Run node --test tests/transcript.test.mjs and node --check workspace.js. Serve with python -m http.server --bind 127.0.0.1. Mechanics tests use synthetic inputs and cover parser validation, source-preserving exports, restoration and text safety. They do not claim transcription quality. The separate local media application is not part of this repository; its earlier checks included file/dependency checks, not verified audio/model processing scenarios.
Load the labeled synthetic example, search for a phrase, select an exact excerpt and export Markdown. The quote keeps its source timestamp. Invalid subtitles are rejected without replacing the previous workspace. This is useful for retaining inspectable evidence from an existing transcript; it does not replace a transcription service or infer meaning.
No customer adoption, performance improvement or commercial usage is established by this demo. Published code is available for inspection; no additional reuse license is granted.
Limits are documented uncalibrated browser resource bounds. Caption markup is preserved as plain text rather than interpreted as HTML. Plain text is segmented by nonempty lines, not semantic understanding. Cue order and wording are retained. Incorrect subtitle timing is rejected. A fingerprint proves text identity, not truth.