Podcast.ai

Inspiration

The All In Podcast features David Sacks frequently saying "I didnt say that" when mentioning a misquote from the New York Times.

What it does

"I DIDN'T SAY THAT" parses youtube to perform topic modeling, and statement indexing that allows you to verify what you've said.

Currently, it's a chrome extension that overlays youtube.

Demo

How we built it

Technical implementation: Youtube transcripts -> Embedding+ Index -> Topic modeling -> Prompt engineering/tuning -> Caching

Deployment: FastAPI to host langchain applications

UXUI: Chrome extension to provide a seamless user experience

Challenges we ran into

1. Collect high-quality transcript data with speaker tagging info for better parsing and summarization.
2. Not enough time, GPUs to train models in one day. We could fine-tune LLM models in parallel to better suit our use case.

Accomplishments that we're proud of

1. Collect and indexing ALL all-in podcast data and embeddings to support various NLP tasks, i.e question answering, semantic search and summarization
2. Prototype the chrome extension to improve the youtube/podcast watching and search experience
3. This could serve as a general fact checker tool for any media content.

What we learned

The default behavior CHANGES EVERYTHING.

What's next for I didnt say that! - the Youtube fact checker

We could run this live across all podcasts and during contentious topics and Presidential debates

Backend API setup

# install python libraries
pip install -r src/requirements.txt

# running the fastAPI server in local
lc-serve deploy local podcast_api

Frontend UI(Extension)

Please refer to this repo