Behind this page is every talk AI Engineer published in 2026, more conference than anyone has time for. Ask any AI engineering question, your agent answers from what was spoken and what was shown, down to the second.
Every still below is a moment your agent can hand back mid-task: the sentence a builder actually said, the text the machine read off the screen behind them, and a link that lands on the second. When a moment has nothing to give, it says so — it never quietly narrows the answer.
heard — every sentence spoken, aligned to the word.
seen — every line that crossed the screen, kept with the box it was read from.
the frame — kept as evidence, even when nothing on it is readable.
Every frame below comes from a talk taken in whole. Point it at a conference, a channel, a creator, and their knowledge compounds into yours.
An agent, on the real protocol, answering a question no single talk answers — with citations, five speakers deep.
call search {"q":"LLM as a judge","content_type":"transcript","max_per_video":1,"limit":6}
← 6 moments · 6 talks · leg: transcript · relevance-first
They agree it is the default and disagree about whether it deserves to be. Two positions, and a third group that ships it anyway:
“And LLM as a judge doesn't really work either because LLMs don't have good taste in writing.”
Nick Heiner · Surge AI · When Will The Benchmaxxing Plague End? · 16:01 · youtu.be/-npY6XjM8CQ?t=961
“Agent as a judge is about adaptive dynamic analysis. LLM as a judge just gives you a fixed rubric with these fixed scores.”
Aparna Dhinakaran · Arize AI · The Future of Evals · 4:07 · youtu.be/q2JrUKBMf0w?t=247
call search {"q":"LLM as a judge in production","content_type":"transcript","max_per_video":1,"limit":6}
← 6 moments · 6 talks · leg: transcript
And three teams who run it in production regardless — video at b_PmGocP4rc (Character.ai), agent evals at 31GUkCBD-Uc (Uber), clinical at O72p-rBb2bA (SonderMind). The disagreement is not “does it work”; it is what it is allowed to be the judge of.
One compose file, one SQLite file, no build step, no runtime network dependency. It runs the CPU half on a Pi.
$ docker compose up -d
clone, copy .env.example to .env, and it is up.
$ claude mcp add --transport http vidtheque https://vidtheque.dev/mcp
one line for Claude, Codex, or anything that speaks the protocol.