What it is
CutPro is a video editor that runs in the browser, made for creators of Reels, TikToks and Shorts. Its main selling point is word-by-word captions for Roman Urdu and Roman Hindi speech, the way people in Pakistan and India actually write online, with English words left in English.
It started from OpenCut, an open-source editor: the timeline, renderer and effects engine came from there. Since July 2026 I've been the main developer turning it into a product.
What I built
- Roman Urdu / Roman Hindi captions: transcription, transliteration, word timing and 35+ caption styles (karaoke, per-word colour and emphasis, animated emoji), plus subtitle export, caption translation that keeps the timings, and a captions-only green/transparent export for Premiere or DaVinci.
- AI editing assistant (OpenAI Agents SDK): 39 tools covering every editor feature. The model runs on the server and the tools run in the browser, so provider keys never reach the browser. One undo reverses a whole AI turn, it can look at frames of the video, and a 14-case eval harness checks it. Users can bring their own OpenAI, Anthropic or Google key (stored encrypted).
- One-press auto-edit: captions, then removing dead air and "ums", then reframing to vertical, then punch-in zooms, with a reason shown for any step it skips.
- Auto-reframe with MediaPipe face and pose tracking, including split screen when two people talk.
- Edit by transcript, Magic Clips (finding the best moments), multicam that cuts to whoever is speaking, B-roll search, chapters and thumbnails.
- Voice tools: voiceover, dubbing of translated subtitles, and zero-shot voice cloning through a self-hosted FastAPI speech service (switched off on the hosted site; it runs when users connect their own speech service).
- About 30 GPU effects written as WGSL shaders in Rust, 3D LUT colour grading and film looks.
- Billing: Stripe checkout, top-ups and customer portal, credits that are held and refunded if a job fails, scheduled reconciliation with Stripe, and the rule that work on your own device is free while cloud work costs credits.
- Shipping: Docker images built in GitHub Actions, staging auto-deploys, self-hosted error tracking and analytics, the marketing homepage and a QA guide.
Hard problems
No engine writes Roman Urdu with word timings. Whisper and AssemblyAI return Urdu script; I also tried Qwen ASR and a fine-tuned Roman Urdu Whisper, but its word timestamps were broken and turning them off made the output fall back to Urdu script and repeat itself. What ships: transcribe in Urdu script, then transliterate with a neural model (86.2% vs 13.8% for plain rules on my test set), with a rule-based fallback that I took from 66.7% to 95.8% exact on the same set.
