Inspiration
We were inspired by Twelve Labs Jockey to create a GPT-4o based tool that has elaborate LLM prompts and an architecture similar to instructor-planner. The fact that the dialogue between interviewers and the speaker is the most important source of information in a press conference inspired us to look for faster transcript text based tools.
What it does
It can extract relevant portions from the whole press conference based on the subjectivity or objectivity of the prompt, allow feedback on generation and share to social media n
How we built it
We built a RAG model for retrieving the most relevant portions and a sanity checker instructor model that removes extracted portions that are not relevant enough.
Challenges we ran into
Hallucinations of the primary RAG model. We solved that by adding a sanity checker LLM to it's output.
Accomplishments that we're proud of
Capturing a range of subjectivity and objectivity of the prompts as well as the feedback system using memory of the LLM conversation.
What we learned
Advanced RAG capabilities
What's next for Samur.ai - AI powered sports press conference summarization
Expanding to more video formats and making the sanity checker more robust by adding memory to it.
Log in or sign up for Devpost to join the conversation.