Status: working prototype. It began as something I built for myself; I’m now working to turn it into something other people can use, not just me. Screenshot shows my own library.
The first thing I ever built was an app for long videos. You give it a ninety-minute talk and it hands you back the two minutes that matter: the key points, the moments worth jumping to, the gist without the sitting-through.
That was the November version: about seven hundred lines of Streamlit, pasted back and forth with ChatGPT before I had ever touched a tool like Claude Code. It was a good toy. It summarized. But I quickly wanted more than a shorter version of the talk.
Two things turned the toy into something I now use every day.
The first: I wanted to ask it questions. Not just read a summary, but actually interrogate what I had watched — what did they say about this, did they back it up, how does it line up with the other talk from last week. So it grew a memory of everything it had read, and the ability to answer across all of it. It stopped being a summarizer and became something closer to a research assistant that had watched the videos for me.
The second is the one I care about most: I wanted to know whether to believe any of it. The people whose videos I was watching were mostly health experts. Long, confident, full of specific claims about sleep and food and exercise. A summary only tells you what the person said. It does not tell you whether they were right.
From “what they said” to “is it true”
So the app grew another job. It reads the video, pulls out the factual claims one by one, and checks each one against the actual research — the same databases scientists use, PubMed and Semantic Scholar and the open medical literature. Not a thumbs up or thumbs down. The evidence: three studies back this up, here they are; this one is disputed, two studies say the opposite. You get to see the ground each claim is standing on.
That checking is the part I care about most, and the part that is hardest to do honestly. Early on, the implementation took the shortest path through an underspecified requirement — I have written about the day I caught it stamping claims as “supported” when all it had really done was match a few keywords. Catching that was what made real checking the point of the whole thing, not a feature on the side.
Where experts disagree
The feature people love is the one that catches the experts contradicting each other. One creator says do this. Another says the opposite. The app notices, puts them side by side, and shows what the research actually says about the disagreement. When two people you trust tell you opposite things, that is exactly the moment you want the evidence on the table.
Underneath, there is a quiet decision that makes the economics work. A video only gets read and checked once, ever. The first person to follow a creator pays the cost of processing them; everyone after gets the result instantly. The knowledge is built once and shared, which is the only way the numbers add up when the expensive part — the AI doing the reading — is almost the entire cost of running the thing. The database, the hosting, the search all round to nothing next to it.
Where it actually stands, to be plain: it started as a tool I built for myself, on my own machine, against my own library. I am now working to turn it into something other people can use. The economics are still the open question — when the reading is almost the whole cost, I am not sure the numbers work for anyone but a heavy user. It works for me. Whether it works as a product for other people is what I am building toward now.
Learnings
I started this as a way to watch less video. It became a way to trust what I watched, and then a way to ask it questions instead of scrubbing through it. The summary was the easy half, and the part everyone else already builds. The hard half — does this claim actually hold up, and can I interrogate the whole library like someone who remembers every video — is the half worth having, and the half that is genuinely difficult to do without lying about it.
The lesson that carried into everything I have built since came from here: the impressive-looking output is the easy part, and the part most likely to be faked. A confident summary is cheap. A claim you have actually checked against the research is expensive, and slow, and the only thing anyone should pay for. Build the expensive half, and be honest when you have not finished building it yet.