SIGNAL EXTRACTION · PRIVATE · LOCAL-ONLY
CH 01 · SIGNAL LOCK
REC · 60HZ · 03:14:22
Learn without watching.
INDEXDIGESTREMEMBER
Learn without watching.
200 AI explainers stacked up. 80% of each one is sponsor read, recap, and "in this video." The actual signal is 5 bullets and 3 timestamps. An LLM extracts that in seconds, for cents.
Drop a Google Takeout CSV. The pipeline does the rest — captions, summarization, tagging, indexing. Every source comes out distilled, browsable, and searchable.
Export Watch Later from Google Takeout. Drag watch_later.csv onto the dropzone. Dedupes against your deck automatically. Test mode caps imports at 10 so you don't torch your API budget.
Each source's transcript is pulled via the captions API. No captions? Paste it manually — the deck won't refuse incomplete signal. Nothing is re-fetched once stored.
Five tiers per source: TL;DR → bullets → notes → quotes. Sonnet writes prose, Haiku tags. Cost: under a cent per source.
Full-text search across titles, all five tiers, and auto-tags via Postgres FTS. Filter by category, channel, date, status. Nothing leaves your deck.
Glance at tier 2 to triage. Read tier 3 to decide if it's worth more. Tier 4 if you want the full thing. Tier 5 when you need to cite the speaker or jump back to the source.
Just the title. The card-grid glance — barely a tier. It's there so the ladder starts at 1.
One sentence, ≤25 words. States what the source actually delivers — not what it promises.
Five concrete claims, ≤20 words each. Skip sponsor reads, intros, recaps, "subscribe" outros.
~300 words, 3–5 short sections with headings. Reads like notes you'd send a colleague.
5–10 quoted moments with click-back timestamps. Cite the actual speaker. Jump back to the moment.
No, and that's not the goal. The goal is to stop watching the 80% of every source that isn't signal. You'll still watch the great 15-minute deep-dive end to end. But for the 40-minute AI explainer where the actual claim is "use schema pinning" — you'll read tier 3 and move on.
Sometimes. Tier 5 quotes are pulled with timestamps so you can verify the source in one click — that's the fact-check layer. Tiers 2–4 are generated; if you spot bad output, hit regenerate.
The deck surfaces them as no captions and gives you a paste-in box for the transcript. About 1 in 40 sources will be caption-less, deleted, or private.
Roughly $0.024 per source. A 200-source backlog is under $5 total to digest. Transcripts use Supadata's free tier for the first 100/month.
Nowhere. v1 is single-operator, self-hosted on Vercel + Postgres. Your archive stays in your DB. No telemetry, no sharing, no public links. Bring your own Anthropic and captions API keys.
Because this is a tool, not a SaaS pitch. The brand is DARKSQUIDINK; the cyan phosphor, scanlines, and courier-log microcopy are structural — they tell you what the system is doing and what it's not.
Drop an operator email. We send a setup guide and the latest deploy notes when the deck is ready for outside testers.