ZENSHO - AI-POWERED DESKTOP VIDEO EDITOR

Zensho is a hybrid desktop application: an Electron/React/Vite frontend drives a locally-spawned Python FastAPI backend over HTTP (bound to 127.0.0.1 for local-only traffic), chosen specifically so CPU/IO-heavy ffmpeg and transcription work never blocks the UI thread. The processing pipeline runs each video through ffmpeg audio extraction, Deepgram Nova-3 word-level transcription, and an LLM rough-cut pass (via LiteLLM, so any OpenAI-compatible or BYOK provider can be swapped in) that decides what to keep, cut, or treat as silence.

Key technical achievements include:

• Custom LCS-based transcript aligner: the LLM returns only the words worth keeping, and a hand-written Longest-Common-Subsequence dynamic-programming aligner maps that refined script back onto the original word-level timestamps, with 'last-occurrence-wins' tie-breaking so repeated takes resolve to the correct (latest) occurrence, and a minimum-LCS-ratio fallback that preserves all content rather than silently discarding it if the LLM paraphrases instead of cutting.
• Waveform-aware boundary snapping: a separate preprocessor snaps every cut point to the nearest word boundary and then refines further by extracting a PCM window via ffmpeg and finding the minimum-RMS-energy point, so cuts fall in natural pauses instead of mid-syllable — with a repair pass that clamps overlaps and removes now-redundant silence segments between adjacent kept clips.
• Serialized async processing queue: a singleton asyncio.Queue worker processes one video at a time through a QUEUED → PREPROCESSING → TRANSCRIBING → ROUGH_CUTTING → COMPLETED state machine, preventing concurrent ffmpeg instances and LLM rate-limit bans while the UI polls a /status endpoint for live progress.
• Non-destructive DaVinci Resolve export: an OpenTimelineIO exporter builds a multi-FPS-safe .otio timeline (plus FCP XML) from the same 'resolved timeline' used for the in-app preview player and MP4 export, linking video/audio clips via matching reel metadata so Resolve treats them as synced pairs — letting editors drag the AI's decisions straight into a professional NLE with every cut still editable.
• Human-in-the-loop editing model: AI-generated cuts (rough_cut_ai) are never overwritten — user edits are stored as a separate rough_cut_human layer, with all exporters resolving rough_cut_human ?? rough_cut_ai per video, so some clips in a project can carry human edits while others still use the raw AI cut.
• Dual interface: the same backend core is exposed both through the Electron desktop app and a standalone CLI (run, process, export, status, config) for scripted or headless use.