Videx

Vision

The best technical talks, deep dives and frontier lectures are on YouTube, and video is the worst
medium there is for finding a specific idea again. Videx turns a library of channels into an index
you can search, quote and interrogate — and it shapes itself around the field you are studying
rather than assuming everyone is here for the same thing.

The problem

Video knowledge sits in unsearchable timelines, spread across creators who never agreed on
vocabulary. A single generic search box treats a shader tutorial and a distributed-consensus talk
the same way, so the queries that would surface the right moment never get suggested and the
summaries never use the right words.

How it works

• Domain personalization — choose AI Engineering, Data Science, System Design, Cybersecurity,
  Creative Tech, or the universal profile. Suggested queries, focus topics, channel discovery and
  AI prompts all retune, and switching is instant.
• Transcript-first ingestion — channel listings via InnerTube, captions, chapters and engagement
  metadata, stored as timestamped segments rather than one flat blob.
• Hybrid search — PostgreSQL full-text for the exact phrase, pgvector embeddings for the idea
  you can only describe.
• Grounded chat — ask a question against one video and every claim links back to the second it
  came from.
• Provider-agnostic AI — a fallback chain across providers, so a rate limit degrades the answer
  rather than the app.