-
-
logo
-
dashboard
-
Project creation
Inspiration
Marketing teams move quickly, but short-form content often loses the qualifiers that make a source claim accurate. “Up to 40% in a 12-customer pilot” can become “Customers save 40%.” We wanted to make that kind of review visible, fast, and grounded in the original recording.
What it does
SourceCut turns recordings into evidence-backed social clip proposals. It keeps every proposed hook, caption, and marketing claim connected to an exact transcript quote and source timestamp. Reviewers can inspect the evidence, select supported clips, render them locally, and hand off videos with their caption, source range, render settings, and evidence package.
How we built it
We built SourceCut with Python, FastAPI, SQLite, FFmpeg, faster-whisper, Kokoro ONNX, GPT-5.6, and Codex. The production desk stores projects, transcripts, clip proposals, render jobs, and outputs locally. GPT-5.6 is an opt-in proposal provider for uploaded projects; every model proposal still passes SourceCut’s deterministic transcript, quote, and timestamp validator before it can be rendered.
Challenges we ran into
The hard part was making AI output inspectable rather than merely plausible. We had to preserve original timeline references through editing, make clip selection end on complete thoughts, keep local CPU rendering responsive, and ensure voiceover and subtitle timing stayed aligned. We also made the no-key Judge Demo reliable without external APIs or third-party media.
Accomplishments that we're proud of
We created a full source-to-handoff workflow: local upload, transcript evidence, reviewer selection, rendering, playable outputs, download actions, and JSON, Markdown, and ZIP handoff packages. Unsupported claims remain blocked even when an AI proposal sounds polished, and the demo works locally with original fictional media and no API key.
What we learned
We learned that trust is a product feature, not just a model feature. Showing the source quote and timestamp at the moment of approval makes an AI workflow easier to review and challenge. GPT-5.6 can propose ideas, while a focused evidence gate decides what can be used.
What's next for SourceCut
Next, we would improve transcript accuracy options, add stronger speaker and scene-aware clip selection, support collaborative review roles, and provide more caption and brand-template controls. The core promise stays unchanged: AI can draft the clip, but every publishable claim must stay tied to its source.
Log in or sign up for Devpost to join the conversation.