LaikaBOT

../LaikaBOT

/Projects/

Armchair Expert AI Podcast Explorer

January 2025

 

The Armchair Expert AI Podcast Explorer was designed to create a fresh and dynamic experience for content exploration. With the advent of GenAI technologies, our ability to search, analyse, summarise, and engage with content has entered an exciting new phase.

 


 


Give it a go here
>> Armchair Expert AI Explorer


What is it?

It's an AI-powered podcast explorer designed to transform how we interact with content. It lets you explore, learn from, and analyse podcast episodes, in a conversational format, beyond the content of the episode.

 

Why is it interesting?

  • You can share your thoughts on the content and get aligned or alternate perspectives through an interactive AI conversation.
  • It allows deeper dives into specific topics for extended learning beyond the content of the episode.
  • For this project, I've used all 830 episodes of the "Armchair Expert" podcast, but the system can easily adapt to other content.

Why is it REALLY interesting?

90% of the build, code and design, was crafted using mainstream GenAI tools, highlighting the power of modern AI to support rapid prototyping and problem-solving. Here's how it came together:

  • Gemini (Google): Kickstarted the project by generating the Python foundation in a Google Colab (Jupyter) notebook. It enabled the app to read episodes from the database, process user queries, and connect with OpenAI's API to leverage ChatGPT-4o-mini.
  • Claude (Anthropic): Stepped in to convert the notebook into a Python web application and HTML user interface.
  • ChatGPT-4o (OpenAI): Assisted with debugging, server-side setup, and filling in where Claude fell short.

I also experimented with OpenAI's Whisper for audio transcription but found faster alternatives for transcribing the 830 episodes — it would've taken a month otherwise!

 

Tips for Building with GenAI Tools

Chunk your data: For this project, I broke each episode into smaller sections and queried each separately. This improved accuracy and reduced hallucinations.

Take it step by step: When using GenAI for developing code, introduce one feature or capability at a time to minimise errors and confusion.

 

It's not perfect, but it works!

The project isn't flawless, and I had to make deliberate efforts to reduce AI hallucinations. That said, the results and the speed at which these tools enabled me to bring the idea to life are something to marvel at… and I've got a great new tool to interact with content for the Armchair Expert and other podcast fans!

 

Have fun exploring the tool!

Give it a go here : https://laikabot.com/armChairExplorer/

 

 











LLM and Gen AI Insights


Context - Whilst LLM's are able to retain large amounts of content in 'context', I found that the more you keep in the open conversation, the less accurate the results became....and in some cases induced hallucinations.

To address this, I developed this application to process transcripts in smaller, manageable chunks. By analyzing these chunks individually and then combining the results, the overall accuracy improves significantly. While hallucinations can still occur, this method greatly minimizes their frequency.

 

The Application Build

 

This web application is 90% built by a blend of technologies from Claude.ai, OpenAI, and Google Gemini.

 

The development journey began with a Python-based Google Colab notebook entirely generated using Gemini technology. From there:
- Claude helped transform the notebook into a functional Python application.

- ChatGPT played a key role in debugging and fine-tuning the experience.

 

For content processing, the app leverages the ChatGPT 4o-mini model.