A podcast episode is not the final output.
It is the starting point for a research process.
20vc.zerve.ai is an independent research layer built on every Thursday Trio episode — the weekly panel with Harry Stebbings, Jason Lemkin, and Rory O'Driscoll.
Every Thursday, autonomous agents fetch the new episode, transcribe it via AWS Transcribe, and attribute every segment of speech to a host. Two pipelines then run in parallel: one measuring how the conversation happens — speaking-time share, topic mix, conversation style, and sentiment per speaker — and one measuring what was said — extracting claims, hot takes, predictions, and open questions via GPT-4o, scoring each for boldness, and synthesising recurring themes across episodes.
The result is a structured, episodic research record of every conversation.
- 01 Fetch RSS feed → identify all Thursday Trio episodes (title, date, description, audio URL).
- 02 Download audio → upload to S3 → diarise via Amazon Transcribe (3-speaker diarization). Resolve spk_0/1/2 → Harry / Jason / Rory via name-mention heuristics (greedy regex).
- 03 Extract claims, hot takes, open questions via GPT-4o — full transcript, agenda-aware topic-boundary chunking.
- 04 Deduplicate claims (string similarity) and questions (semantic embeddings + cosine similarity). Verify open questions — resolution pass removes answered ones. Tag recurring cross-episode questions via semantic clustering.
- 05 Score claims for boldness algorithmically (specificity + directional strength + contrarian signal + stakes → 1–10); hot take = ≥ 7.
- 06 Compute speaking-time share per host (Harry / Jason / Rory) per episode.
- 07 Classify topic mix — 7 clusters (AI, Venture Capital, Future of Work, SaaS & Enterprise, Private Markets, Macro / Markets, Policy) via keyword-cluster scoring; normalised to % by episode.
- 08 Classify conversation type — 6 types (Visionary, Feisty, Analytical, Contrarian, Reflective, Tactical) via signal-capped ensemble classifier on topic-shift-segmented conversation blocks; time-weighted by block duration.
- 09 Score sentiment — VADER on natural conversation blocks (not sentence-level); duration-weighted pos / neu / neg per episode; split by intra-block word-count share per speaker.
- 10 Synthesise open questions via GPT-4o — deduplicate globally, rewrite concisely, score report-worthiness, map back to episodes.
- 11 Generate episode summaries via GPT-4o — punchy 2-sentence website teasers.