The research layer for 20VC
Every episode analysed. Every hot take challenged.
Every open question investigated.
Latest Episode
Wall St’s $725B AI Question
Every claim checked. Every number traced back to source.
Across every analysed episode
4 episodes · 232 minutes · 33,600 words transcribed, diarised, and classified.
Who's talking
DiarisationConversational vibe
Zero-shot classifierTopic mix
7 topicsSentiment
RoBERTa fine-tunedFor the geeks
Diarization
Multi-speaker embedding extraction using x-vector architecture. Employs spectral clustering with PLDA scoring for robust speaker identity separation. Error rates (DER) optimized for overlapping speech segments and varying acoustic environments using neural voice activity detection (VAD).
Topic Mix
Unsupervised Latent Dirichlet Allocation (LDA) models the statistical distribution of latent themes. Dirichlet priors are dynamically adjusted based on corpus volume to maintain thematic density. Hierarchical topic modeling reveals sub-cluster dependencies with high coherence scores.
Conversation Type
Dialogue act classification via bidirectional transformers (BERT). Maps utterance intent across informative, directive, commissive, and expressive categories. Finite State Machine (FSM) tracking maintains contextual state across multi-turn asynchronous exchanges.
Sentiment
Aspect-based sentiment analysis (ABSA) utilizes attention-weighted recurrent neural networks to detect polarity at the entity-attribute level. Sentiment shifts are tracked with millisecond-precision timestamps, mapping emotional trajectories against specific conversational milestones.
A podcast episode is not the final output
How AI Co-Research agents run a full analysis every week, automatically