VidScope

Fair point, and you've named the real weak spot. Today VidScope works from the transcript only: no frame OCR and no matching against a package index. So when a name is only on a slide or in the code and the speaker says "this one", it's missed. When the ASR mangles a name, it's only caught if the surrounding spoken context is clear enough. That's the trade-off I made. I'd rather miss a tool than invent one, so the 0% claim is about precision. It says nothing about recall, and I haven't measured recall on conference talks. Slide OCR would be theFair point, and you've named the real weak spot. Today VidScope works from the transcript only: no frame OCR and no matching against a package index. So when a name is only on a slide or in the code and the speaker says "this one", it's missed. When the ASR mangles a name, it's only caught if the surrounding spoken context is clear enough. That's the trade-off I made. I'd rather miss a tool than invent one, so the 0% claim is about precision. It says nothing about recall, and I haven't measured recall on conference talks. Slide OCR would be the natural next step for that kind of video, and your 45-minute talk example is a good test case. Thanks for the sharp question. natural next step for that kind of video, and your 45-minute talk example is a good test case. Thanks for the sharp question.

Names are also checked against the transcript, so a mangled name the model can't match to the text gets dropped rather than guessed.

Yes. In the web app, each detected tool shows the timestamp, a verbatim quote from the transcript, and a "watch in video" link that opens YouTube at that exact second. Quotes are checked against the transcript, and anything that isn't verbatim gets dropped rather than shown. The downloadable Markdown/PDF report currently only has timestamps in the "key moments" section, not per tool. I'll add per-tool timestamps and quotes there too.

Thanks Andrew, fair questions. Today VidScope works from YouTube's captions (manual when they exist, auto-generated otherwise). We don't re-transcribe the audio yet, so you're right that garbled tool names are a real risk. What limits the damage is that the model extracts tools using the surrounding context, not just the spelling, and it only keeps what the speaker actually said. Re-transcribing with a stronger speech model for auto-caption-only talks is on my list, and your comment moves it up. On links: they're not from a curated index. Each tool is searched with its category and description as context, then the page is checked. The name has to appear in the domain, path or title, and the page has to look like software, not just contain the word. GitHub repos are accepted only as exact owner/repo matches. If confidence is too low, no link is returned. A missing link is better than a wrong one. I had exactly this problem with names like Neon, Codex and Serena, which is why the validation is strict now.

Thanks!

Thanks Zhongcheng, good question. For a name with no link in the description, VidScope uses what the speaker said around it (category, what the tool does) as search context. That disambiguates most cases, like database "Neon" versus the unrelated sites that come up for the bare word. The candidate page then has to match the name and show tech signals. For GitHub-only tools, it requires an exact repo-name match. The limit is that grounding stops at the transcript. For internal codenames or two libraries with near-identical names where the speaker gives no context, there's nothing to ground on. In that case VidScope returns the tool without a link instead of guessing. I'm considering marking those as "unverified" so you can see where it wasn't sure.

Thanks Juan, that's exactly the pain point I built it for. Strict grounding in the transcript was a deliberate trade-off: I'd rather miss a vague mention than invent a tool that was never said. Your "that tool from X company" case is a real gap. Flagging vague mentions for manual review, rather than guessing, fits that philosophy well, so I'm adding it to the list. On docs links: each extracted tool already comes with its official link, so that part is covered. If you try it on a conference talk, I'd love to hear which mentions it missed.

Thank you!

Thanks! 🙏

That's it!

Thanks, that's exactly the use case. "10 tools you should know" videos are where names get lost fastest. Yes, VidScope includes timestamps: each tool comes with the moment it's mentioned, so you can jump back to that part of the video. You also get the official link and the Markdown export, so you can save the tools worth trying. If you test it on a video and something's off, tell me and I'll look into it.

Thanks! 🙏

Thanks! 🙏

Thanks! Strict grounding is the rule: a tool is only listed if it appears in the transcript. When a name is mangled by the captions, we'd rather leave it out or flag it as unresolved than guess a package name. Official links are only attached when the name resolves confidently. Handling near-misses better is on our list, and it's exactly the kind of feedback that helps prioritize.

Thanks! 🙏

Thanks! 🙏

Thanks!

Great point, an audit-friendly evidence section is the right direction. Yes, each tool comes with its timestamp. Timestamp + sentence per tool is a great idea, noting it for the next iteration

Thank you, really appreciate it! 🙏

That's a really useful suggestion. Grouping repeated mentions under one official URL and putting unresolved names in a separate section would fit our transcript-grounded approach well, and it would make the export a real research checklist. Adding it to the roadmap, thank you!