Something dumb about AI agents that bugged me enough to go build the fix: send one off to research "what does X actually say about Y" and it comes back citing SEO listicles and blog recaps, half of them AI-written summaries of other AI-written summaries. Meanwhile the actual answer is sitting in some 3-hour podcast the agent literally cannot hear.
The DIY route is miserable. I know because that's what I've been buried in: pulling audio, transcription, speaker diarization, chunking, embeddings, vector DB, and keeping the whole thing fed as new episodes drop. It's a full-time infrastructure job disguised as a feature.
So I built Jamie (pullthatupjamie.ai) to be that infrastructure so nobody else has to run it. Semantic search over 150k+ indexed hours of podcast audio. One API call returns the exact moment a topic came up: quote, timestamp, speaker, playable clip. Receipts instead of recaps.
The part stackers will care about: ~$0.10 a call, no subscription, no account, takes Lightning. Your agent can pay with L402 for its own credits when it runs dry. Machine paying machine for information, v4v style, working right now, not a whitepaper.
If you want to kick the tires, point your agent at https://www.pullthatupjamie.ai/for-agents (it can read the docs itself), or if you're on PPQ it's already a plugin: https://ppq.ai/models/podcast-search
Happy to answer anything about the stack, the indexing, or the Lightning flow. And curious what everyone else is feeding their agents for research.
The other day I was trying to get some Jeff Booth clips but instead I got Saylor clips, and aantonop @ Joe Rogan. Is there a way to filter in the query, i.e:
not:"Joe Rogan" definitelynot:"Saylor"?no determinstic filters like that but you can phrase a similar idea in natural language and it's usally good at applying those.
I.e. "Find me Jeff Booth on X pod or within the last 3 months" etc. It picks up on feed filters, time range filters and topic filters from the natural language.
@optimism FWIW I also learned that agents are really bad at following negative instrucitons. I.e if you say "don't think of a pink elephant" it will immediately do that lol
Yeah. Context poisoning! That's why the golden rule is to never argue and just do it again.
i actually think there's a human parallel. There's noise when you make it focus on negative outcomes. I noticed this when I'm coaching I'm always leading with positivity and only going negative if they're really not trying and need a kick in the ass.