We extract user interests from digital exhaust (CRM entries, web logs, and emails) using embeddings -- a low-cost way of using LLMs.
We use this to name the interests using the prompt Given this information about a financial advisor, return a JSON object {\"interests\": \"...\"} containing comma-separated keywords they would be interested in when listening to a financial podcast.
We transcribe audio files using the open-source Whisper model. It's great at stripping out "uh...", "um...", etc. and creating a cohesive transcript.
We don't diarize the transcript to add speaker names, but using Anthropic's Claude 3 Haiku,
a frontier model, we can do this with a simple prompt:
Label this call transcript between an agent and caller based on context
.
We create structured summaries using the prompt:
Given this video transcript: Title: ${title}. Transcript: ${transcript_with_timing}. Write a few paragraphs for: ${profile}. Explain the implications for their investors and how they can leverage this information to maximize returns for their clients. Respond as JSON paragraphs with timing.
The entire analysis for a 20 minute video takes under half a minute.
For Whisper on OpenAI with Claude 3 Haiku, the typical cost of a 20-minute video personalized for 6 advisors is under 30 cents, and grows at about 3 cents per advisor.
Yes. With open-weight models like Whisper and LLama-3, you can run this in your data center (or even your gaming laptop).
This is not a product. It's a demo. Email [email protected] for details.