You cover a company, and three streams of information reach you: the filings, the numbers, and the executives. The first two are structured, audited, and priced in within minutes. The third is the loosest — and often the earliest.

Executives talk constantly: earnings interviews, conference panels, long-form podcasts, keynote Q&As. Spoken language leaks what documents are drafted to hide — a CEO who stops saying "we will" and starts saying "we aim to", who answers a margin question with a story about culture, or who quietly reassigns credit and blame between quarters. That record is sitting on YouTube right now. Assemble everything one executive has said in public and read it the way you read a filing — line by line, source by source, over time.

deep_search does the assembling. Ask for a person's interviews and it runs a transcript-aware sweep of YouTube — not just the most-viewed clips. You can restrict results to actual interview, podcast or news content, limit them to a recency window (this week, this month, this year), sort newest-first, and prefer full-length conversations over shorts. Each result card comes back with transcript-derived labels, an evidence snippet with a timestamp pointing at the exact moment your query is addressed, and quality verdicts — so a low-effort reaction video never enters your record by accident.

deep_query does the reading. Ask a question across that record — "how has their language about margins changed?" — and it runs the full research pipeline over the transcripts, returning a grounded answer where every claim carries an inline citation to the second. You are never trusting a summary you cannot check: every sentence links back to a moment you can watch.

Both tools run inside the AI assistant you already use — Claude, ChatGPT, Cursor, or any MCP client — so the whole workflow is a conversation, not a research project.