From Bilibili and TikTok to one readable digest

The problem

Follow a few creators and the backlog grows by the dozen. Finding one specific point of view means watching everything, and remembering none of it.

The hard part is not downloading. It is deciding which video is worth watching — platform recommendations are optimised for keeping you there, not for answering your question.

What we built

A four-stage pipeline where every stage can be re-run on its own:

1. Topic filter — match against the themes, keywords and creators a user actually cares about, and drop everything else before it costs anything downstream.
2. Fetch — store the video plus metadata, keeping the original link for traceability.
3. Transcribe — run WhisperX on-premise to produce a timestamped transcript. Nothing goes to a third-party speech API.
4. Summarise — feed the whole transcript set to a large model and ask for what the videos are saying, which points repeat, and where they contradict each other.

Three decisions worth explaining

Why transcription stays on-premise. A transcript is the full text of the video, and it often contains commercial information that has not been published yet. Sending that to a cloud speech API is not something a content team can accept.

It also turns out to be cheaper: per-minute speech APIs add up quickly, and at volume the on-premise run wins. This time security and cost happened to point the same way.

Why not summarise each video separately. A single-video summary has limited value. The useful signal is cross-video: across ten videos, which claims repeat, and which ones disagree. That only works when the text is collected in one place.

Why keep timestamps. A summary gives conclusions, but people need to verify them. With timestamps, an interesting paragraph links straight back to the source.

The outcome

The shape of the work changed: from "how many videos do I have to watch today" to "read the digest, click back into whatever sounds interesting."

We deliberately did not build "automatic quality judgement". It is not reliable yet, and shipping it would cost the user's trust. The pipeline does what it can actually do: transcribe, deduplicate, extract.

What is not done

• No incremental updates — it runs in batches
• No cross-language translation; transcripts stay in the source language
• Summary quality depends on prompting plus spot checks; there is no systematic evaluation yet