From raw audio to decision-ready intelligence, at scale.
We capture the spoken word across podcasts and news broadcasts — local stations and national programs alike — transcribe it with a machine-learning pipeline, and let agentic AI turn millions of minutes of audio into structured, trackable intelligence.
Live pipeline figures, updated continuously.
How the engine works
Three stages, fully automated — from the moment an episode publishes to a structured insight a strategist can act on.
We ingest podcasts and news broadcasts at scale
A fleet of fetchers continuously pulls new episodes the moment they publish — podcasts alongside local and national news broadcasts, from hometown stations in every market to the biggest national programs. Audio is downloaded, deduplicated and queued automatically.
A machine-learning pipeline turns audio into precise transcripts
Every recording runs through high-accuracy speech recognition, producing word-level, time-coded transcripts. Advertising is automatically stripped and each segment is structured, so downstream AI reads clean, editorial content — not filler.
Agentic AI reads every transcript and extracts structured intelligence
Autonomous agents work over the transcript corpus — recognizing entities and cleaning them into canonical identities, clustering stories, scoring sentiment and framing, and writing week-over-week narratives. The result isn't keyword hits; it's a living, structured picture of who is being covered, how, and where it's moving.
A stack of purpose-built models
Each layer is a dedicated model or agent. Together they convert unstructured speech into a queryable intelligence graph.
Speech recognition
High-accuracy transcription tuned for the messiness of real spoken-word news audio.
Named-entity recognition
A recognition model surfaces every person, organization and place spoken across the corpus.
Agentic entity cleanup
AI agents resolve messy mentions — “Trump,” “President Trump,” “Donald J. Trump” — into one canonical identity, and prune noise the model isn't sure about.
Topic discovery
Semantic embeddings are clustered into evolving story threads, with no keyword list required.
Agentic insight generation
AI agents name clusters, write narratives and track how a story drifts week to week.
Sentiment & framing
Per-segment tone and framing scoring, calibrated for news coverage of public figures.
Left / center / right divergence
Coverage split by outlet leaning — the contrast keyword tools can't see.
Drift & momentum
Week-over-week movement scoring flags stories accelerating right now.
Not keyword alerts. Structured intelligence.
Legacy monitors tell you a word was said. Catchwind tells you who was covered, how they were framed, whether the left and right diverged, and which way the narrative is moving — computed across the entire audio corpus, every week.
See the engine at work
Explore live intelligence built from the pipeline — or talk to us about your coverage.