MiVid

The problem

Cutting a podcast into social clips takes hours of finding moments, reframing, captioning and styling each one.

What I built

• AI finds the most shareable moments in a long podcast and cuts them out
• Smart Reframe: converts landscape to vertical by detecting shots and following faces, including two-speaker shots
• Word-by-word transcription and animated kinetic captions that avoid covering faces
• Large style library: 1,800+ fonts, 56 animations, 30 text styles, one-click themes and color grades
• Caption templates learned from a screenshot of a style I like
• Motivational mode: upload a voiceover and get a full video, with backgrounds from a B-roll library or 16 built-in animated scenes chosen to match the words
• Background music mixed automatically at the right volume
• A web app to upload, pick a style, watch progress and download clips

How it's built

A Python pipeline (ffmpeg, Whisper transcription, face detection, Gemini) that resolves every caption decision, and Remotion (React) that renders each frame of the final video.

• Python
• Remotion (React/TypeScript)
• faster-whisper
• Google Gemini
• FastAPI
• ffmpeg
• OpenCV

My role · Designed, built, deployed and maintained solo with AI-assisted development (Claude Code).

Proof

• Two complete modes (podcast clips and voiceover videos)
• Every render is deterministic: the same input and settings produce the identical video

Hiring for a leadership role? Book an interview