The experts are already talking. You’re not in the room.

Every week, the people who actually know their field, in venture, healthcare, energy, defence and public policy, talk for hundreds of hours on podcasts. PodInsightHQ’s AI listens to all of it, finds where they agree and where they disagree, and lets you hear them say it in their own words.

8,100: episodes
13: themes
7,300: hours transcribed
1,200,000: mentions of companies, people and products

Every episode published since January 2025, growing by around 100 episodes a week. We started with technology and venture, and are now adding healthcare, energy, markets, defence, brands and public policy.

Built by two founders directing AI agents, for under $3,000 of cloud credit. It runs under $3 a day when idle.

Why it matters

Nobody can listen to it all: One theme produces over a hundred hours of expert conversation a week. Thirteen themes is a full-time job for a room of analysts.
The useful part is in hour two: Long-form conversation is where practitioners think out loud. It is exactly what headlines and summaries strip out.
Summaries can lie. Audio can’t: Every claim we surface plays back the original words. If we cannot play it, we do not publish it.

How it works

Listen: Every episode is transcribed word for word, with a timestamp on every sentence.
Understand: AI reads every transcript and pulls out what is actually being claimed, and who and what is named. Different spellings of the same company are reconciled so it counts once.
Connect: One expert saying something is an opinion. Three saying it independently, on different shows, is a pattern. We find those, and we find the disagreements too.
Prove: Every claim carries its episode, speaker and timecode. Press play and the clip is cut from the original audio at that moment. A claim that fails this check is dropped, not published.

Built on AWS

The whole system runs in AWS in Frankfurt. It costs close to nothing when idle and scales up on ingestion days.

Speech to text: Amazon Transcribe converts every episode to timed text.
Search: Amazon OpenSearch Serverless holds the vector index, with embedding and reranking models run through Amazon Bedrock.
Language models: Amazon Bedrock running Claude for claim extraction, entity extraction and written briefs.
Pipeline and storage: AWS Lambda and Amazon S3 handle ingestion, audio clipping and raw storage in eu-central-1.

Where we are

PodInsightHQ is a private build. The system is running against the full library, and the intelligence layer is in active development. It is not open for sign-up, and there is nothing to buy yet. If you want to talk about it, email james@podinsighthq.com.