Context
I inherited a mock that generated commentary from a placeholder transcript, not from the video itself.
Approach
- Built a real pipeline: a clip goes to Gemini (Flash or Pro), commentary comes back in a selectable style with timed SRT subtitles, and FFmpeg burns them into the video.
- Split the backend into LLM, prompt and video-processing layers, and showed processing time in the UI so the two models could be compared.
Outcome
A working end-to-end demo in about 2.5 weeks.