The Frozen Asset
How fifteen years of conference video became something an organization could actually act on.
By Felipe Lujan-Bear · LiquidBook
For fifteen years, Agile Alliance recorded its conferences. Keynotes, panels, workshops — the accumulated thinking of a global movement. More than four hundred talks in all, over three hundred hours of video, gated behind a login and hosted quietly on Vimeo.
Somewhere in that archive was gold: ideas as sharp today as the day they were spoken. And somewhere in the same archive was rot — talks anchored to tools that no longer exist, predictions about years already gone, case studies from companies that have since folded. Institutional knowledge and dead weight, filed side by side, indistinguishable from the outside.
The problem was never storage. It was knowing what they had. And the only honest way to find out was to watch all three hundred and ten hours.
No one was ever going to watch three hundred and ten hours.
Every organization that has produced content for a decade is sitting on some version of this: an asset it can no longer see the shape of.
What we actually did
So we didn’t watch them.
Over the course of about a week, a pipeline read the audio of every talk, transcribed it, and weighed each one against a single question: is this still worth a viewer’s time — and if not, why not?
Each of the four hundred and thirteen videos it processed came back with one clear recommendation — keep it, keep it but flag it as dated, or review it for retirement — paired with a two-or-three-sentence explanation in plain English. Not a confidence score buried in a dashboard. A reason a staff member could read, agree or disagree with, and act on immediately, with no training and no new tools to learn.
The whole thing landed as a single spreadsheet. Sort by recommendation, click the link, decide.
The number — and what it actually means
Here is the part that tends to stop executives mid-sentence: the AI analysis, across all four hundred and thirteen videos, cost about one dollar.
I want to be precise about that figure, because it is easy to misread. It is not the price of the engagement, and it is not an offer. There is real cost behind work like this — the workstation it ran on is a four-thousand-dollar machine, and skilled time went into designing, building, and supervising the pipeline. The dollar is not what the project costs.
The dollar is evidence of something larger. The operational cost of pointing serious AI at a real problem has collapsed. Work that a few years ago would have demanded cloud infrastructure and a line-item budget now runs — for the compute itself — on pocket change. That shift is the actual story. Not “we’ll sort your library for the price of a coffee,” but “the economics that used to make a project like this hard to justify have quietly evaporated.”
“We were downloading seven-hundred-megabyte files just to throw the video away.”
A mid-build discovery: transcription only ever needed the audio. Catching it cut the download footprint by nearly ninety percent — the kind of thing you find when a person is actually watching the machine work.
Recommendations, not decisions
The moment you tell an executive you turned AI loose on fifteen years of the organization’s knowledge, they tense. And they should. So the architecture was built to earn trust before asking for it.
The machine recommended. It never decided. Nothing was deleted, nothing was auto-actioned, nothing changed unless a human chose it. The entire process ran against a safe staging clone of the site and never touched the live system. And of all four hundred and thirteen videos, not one came back as “I can’t tell” — every single one carried enough signal for a confident call, which staff remained free to accept or overrule.
That last point matters more than it first appears. The value was never the automation. It was that the pipeline did the discovery — the tedious, impossible part — and handed the judgment back to the people whose judgment it should always have been.
The most important design decision on the entire project was deciding what not to hand to the machine.
Built into what they already owned
This is where a project like this either stays a party trick or becomes infrastructure.
The pipeline was not a standalone gadget with its own login and its own silo. It reached into the WordPress system Agile Alliance already ran, spoke its native structure, and dropped staff back into the admin they already knew — one click from a row in the spreadsheet to the exact edit screen for that video. No migration. No new platform to adopt. The AI met the team inside the tools they already used every day.
And because it genuinely read the library rather than guessing at it, it surfaced something the organization had lost track of about itself. The archive isn’t only in English. Thirty-five talks are in Spanish, twenty in Portuguese — an entire multilingual dimension, drawn from years of Latin American and Brazilian conferences, sitting unmapped for over a decade. Nobody configured that. The pipeline detected each language on its own and handled it in stride.
The system understood their library better than they did.
What comes next is a decision, not a build
Here is the quiet punchline.
Every transcript the pipeline produced already exists. And because those transcripts live inside the WordPress system Agile Alliance owns, that same text can — whenever the organization chooses — flow back in and make fifteen years of talks fully searchable on their own site for the first time. Someone looking for “psychological safety” would surface every talk that ever touched it. The capability is built. It is waiting.
That is usually where a project like this truly lives or dies. In my experience, the technology is rarely the hard part anymore. It’s ready, it’s proven, and it is astonishingly cheap to run. What remains is the decision to reach for what you already have.
And that is the real question this kind of work sets in front of an organization. Not can it be done. Not what will it cost. Not is the technology there yet. All of that is settled. The only question left is whether you pick it up.
For the technically curious
The spine above is the story. This is the machine room, for anyone who wants to see it.
- Audio-only ingestion. Full videos ran 200–700 MB each; transcription never needs the picture. Switching to audio-only streams cut the download footprint by roughly 87%.
- Local transcription, zero API cost.
faster-whisperon an RTX-class GPU handled a 73-minute talk in about 75 seconds — 310 hours of audio transcribed for nothing but electricity. - A cheap, capable judgment layer. A lightweight model (GPT-4o mini) condensed each transcript and generated its recommendation for roughly $0.002 per video.
- Fully resumable. Every video’s progress lived in a SQLite queue. A crash, a power cut, a scheduled reboot — the run picked up exactly where it left off, with no reprocessing.
- An assembly-line “wave” design. Processing every video through one stage before advancing solved a stubborn GPU-cleanup crash by isolating transcription in its own subprocess — the parent read committed state and carried on.
- Built and supervised live with an AI coding agent. Which is precisely why the wasteful-download insight surfaced mid-session, as a fix, rather than after the fact as a regret.
LiquidBook — specialized WordPress, integrated with AI, built into the platform you already own.