Looking for Advice: Turning Meetup Videos into an AI Product Mentor

AI Discussions
,
1

I run a monthly meetup with ~100 product leaders. We record each session, and I’d love to turn these transcripts into an AI “Product Coach” that can answer questions grounded in our discussions.

My initial ideas:

  • Chunk & embed transcript segments into a vector DB (thinking Pinecone or Weaviate)

  • Summarize every 5 minutes into semi‑structured “cards” before embedding

  • Auto‑generate FAQs from the transcript with synthetic Q&A pairs

  • Build a knowledge graph to connect related concepts and speakers

My goals:

  • Let members ask targeted questions, such as “What did Jane Smith say about product-market fit in March?”

  • Keep answers accurate and grounded in the transcript

My questions:

  1. Best practices for chunking & embedding long transcripts?

  2. How to generate FAQs without introducing hallucinations?

  3. Has anyone combined summarization + knowledge graphs for event content?

Any insights, tools, or examples would be hugely appreciated!

2

Just adding a bit more context in case it helps spark ideas:

I’m aiming for something lightweight enough to run without a huge infra budget, but still accurate enough that people trust the answers. I’m open to any stack — LangChain, LlamaIndex, RAG pipelines, you name it.

Even if you’ve only tried part of this (e.g., transcript chunking, FAQ generation, or knowledge graphs), I’d love to hear what worked and what didn’t.

What’s one tip or lesson you’ve learned from building an agent like this?

3

break transcripts into speaker chunks with timestamps, then also make short 5-minute summaries as “cards” to embed. When answering, always show the exact quote and time so people can check it. Only keep auto-made FAQs if the transcript really supports the answer. A small graph linking people and topics can help search, but don’t overcomplicate it. Biggest lesson: trust comes when every answer links back to the original moment.
lets connect
I might be able to help ya.

1 Like
4

have you tried using crewai agents?

It surely does most of the work you mentioned except knowledge graph which I suppose you can include if you know how it’s used.

1 Like
5

Thanks for your comments. To create a good auto-FAQ maybe I could first list all questions, try to answer grounded on retrieved content, and only keep what has a grounded answer.