About MFMVault.com
MFMVault.com is a fan site made by Greg of Leverage.
This is a proof of concept for a more general purpose podcast player and vault.
The hypothesis is that to better engage with an audience, highly currated, ai-enabled (but human verified) data is key.
Where AI was used:
- Transcript formatting: You'll notice on a segment page, the transcript is formatted (see the bold, quotes, and nice dollar signs?). The transcript out of the transcription service does not have this. We used AI to interpret the text and add the pretty transcript formatting to it. After prompt engineering, we used gpt-4o due to cost and performance.
- General Coding: Via Cursor
- UI: Heavy influence from v0.dev
- Insights: All insights (Frameworks, Quotes, Products, etc.) were derived from the transcript using AI. Not only to extract the insight, but also identify when the people talking about the insight started and stopped. See a summary of this process here.
- Segment Titles: Use gpt-4o to generate a title for the segment. This is based off the transcript or content
- Research: To hydrate the guests with twitter profiles, bios, and pictures, we used Perplexity.ai to take the first pass at sourcing information. Then it was human reviewed.
- Chat Bot: While there are many options for chat bots, we used OpenAI's assistants to get to market quickly. The speed-to-product is hard to beat. This will need optimizing later.
- Related Insights: Each insights page has a "Related Insights" section. This uses text embeddings to find other insights that are related.
Our philosphy is that AI will speed boost us, but we still need humans to verify the data in most cases.
Tech stack:
- Next.js - Frontend framework
- TailwindCSS - Styling
- Radix UI - Component library
- ShadCN - Component library (heavily inspired the UI)
- Lucide Icons - Icons
- Deepgram - Transcription
- Supabase - Database & Storage
- Redis / Celery - Background Jobs
- Railway - Hosting
- Flask - API
- Perplexity - Research
- Exa.ai - Research
- Cursor - Coding
- Meilisearch - Search
- OpenAI Assistants - Chat Bot
- LangChain - For orchestration
- LangGraph - For observability
- OpenAI - Main models and assistant API
- Anthropic - Claude 3.5 Sonnet as a workhorse
Special thanks:
- Arvid for inspiring me to try out Meilisearch and UI inspiration via Podscan
- Pat for inspiration on UI with StarterStory
- Finn for helping with data collection and curation