what is trimora_v2 fr?

Trimora takes a long video, like a podcast, lecture, or interview, and automatically finds and cuts out the best short clips from it, the kind of 45 to 90 second vertical clips people post on TikTok, Instagram Reels, and YouTube Shorts. Instead of a person scrubbing through hours of footage looking for good moments, Trimora does that work automatically using a mix of rule based scoring and AI language models. Under the hood, the project is built as a pipeline, meaning the video passes through a series of steps in order. First the audio is transcribed to text, then the transcript is analyzed to understand its meaning and structure, then candidate clips are generated and scored, and finally the best set of clips is chosen and assembled. The README describes this pipeline as production grade, meaning each step is designed to behave predictably and fail with a clear error message rather than silently producing bad results. One of the more distinctive parts of the system is what it calls the deterministic story composer. Rather than asking an AI model to guess where a good clip should start and end, this component uses a search technique called beam search to try out many possible edits and score them along three measures: whether a segment fits the same story, whether it adds value like a surprising or emotional moment, and whether it helps bring the story to a satisfying end. The weight given to each of these three measures shifts depending on whether the segment is near the beginning, middle, or end of the story. The project is built with Python and FastAPI on the backend, React for the frontend, and FFmpeg for video processing. It also uses Faster Whisper for transcribing speech, with the option to fall back to a cloud based transcription service when a local GPU is not available. According to the README's own time estimates, transcribing and processing an 8 minute video takes roughly 25 to 40 seconds with GPU transcription, while a 1 hour video takes about 2 to 4 minutes. The project is licensed under MIT, which is a permissive license that allows free use including commercial use. The full README is longer than what was shown.