Hey folks! 🎧 I had the most enlightening conversation on the latest episode of Data Chaos. We delved deep into the maze of coding with Sergio Leon, CTO extraordinaire who has traversed the fields of AI and Signal Processing with a dashing Cuban flair. 💻🌴
We navigated his 17-year journey from his college days creating desktop applications to breaking new ground in the world of AI. Ever heard of a mobile app startup in Cuba running on email attachments? Well, Serge and his team did it! 💡🚀Oh, and did I mention we also explored the vast realm of AI? Serge's idea for improving AI efficiency through feature selection in high dimensional vector space was nothing short of groundbreaking. 🌌🧠And of course, we took some time to appreciate YOU, our listeners! Because what's a journey without a community to share it with, right? 🙌
Tune in to hear about Serge's Cuban roots, software innovations, and all things AI. Trust me, you don't want to miss this one. 🎙️💥
(00:02) GraphQL Evolution in Software Development
Serge, a software engineer with 17 years of experience who grew up in Havana, Cuba and moved to Miami when he was 26, shares his journey into software engineering. He started in college by building a desktop application with Visual Basic and graduated from an electrical engineering curriculum but added computer science to it. We also discuss his experience with GraphQL and how Propel decided to go all-in on it, benefiting from the ability to inject filters and get complex data back in a single round trip.
(08:26) Mobile App Startup in Cuba
Serge talks about his experience with mobile app startup KIPaoi in Cuba. Despite the lack of mobile data at the time, he and his team developed a content distribution network using email attachments to send data. With their platform, they were able to generate 8,000 monthly active users and even provided a custom version of the app for the International Book Fair of Havana. Serge shares his story of how he and his team of Cuban software engineers created their innovative mobile app, and he even recounts the time President Obama welcomed them during his keynote speech at the Global Entrepreneurship Summit.
(19:05) AI and Tech Hub in Miami
Serge shares his experience transitioning to the software engineering world and his journey to AI. His entrepreneurial spirit was sparked from a young age when he made paper planes to sell. He talks about his EE degree and how he went into data analytics and IoT. He chose to move to Miami and the advantages it has in the technology industry. The lockdown has demonstrated that you can build anywhere in the world with the right talent and drive.
(31:31) Tech Stack and Developer Experience Discussion
Our small team of developers has focused on optimizing for user experience with Web Transitions API, Veet, and Nodejs. We have kept our tech stack simple by using container for our UI management and Docker swarm for our deployments. I have tried to build with cost-efficiency in mind, and have chosen declarative frameworks such as CDK over Terraform to help with repeatability and infrastructure.
(47:17) Language Model Compression & Vector Dimensionality
Serge discusses his idea to improve the efficiency of AI and LLM's by using feature selection based on entropy in a high dimensional vector space. He explains the process of taking customer knowledge base data, calculating embeddings, and using those embeddings to generate questions and answers. He also outlines the problems with using logarithmic search and how he hopes to increase efficiency with his feature selection technique.
(59:14) Improve Propel's SDL Organization and Retrieval
We explore how to make AI and LLMs more efficient by using feature selection based on entropy in a high dimensional vector space. Serge talks about the use of LangChain and Lama Index to vectorize, chunk and index data, as well as the use of Propel Data API schema to write a time series query. We also look into the debug output to understand how the LLM makes choices, and the potential of inferring questions and types of data to present to customers. Finally, we discuss how to leverage comments from the customer and the strong definition of the metric types available to aid in the selection of the specific segment of the SDL.
(01:06:33) Improving Vector Search and Choosing Database
We discuss how to optimize user experience with Web Transitions API, Veet, and Nodejs with a small team of developers. Serge shares his idea to improve the efficiency of AI and LLMs by using feature selection based on entropy in a high dimensional vector space. We explore the trade-off between dropping dimensions and the accuracy and performance of the output, and why Jerry the founder of Lama Index decided to create his own platform instead of joining Langchain. We also discuss how utilizing RSDL and R data led Serge to Lama Index, and why it's easier to start using Lama Index for augmented generation.
(01:20:22) AI and Rapid Application Development Discussion
We explore the potential of using AI to rapidly develop applications and data access layers. We discuss how React Admin, a repo built on top of Material UI, uses data providers to auto-generate React components with Material UI. We consider how AI can be used to refine this process. We also agree to a follow-up conversation in a month or so to see where our AI and LLM journey has taken us.
(01:25:11) Word-of-Mouth and Live Stream Appreciation
Propel empowers its customers with data and the results they are getting. AI can be used to rapidly develop applications and data access layers, with React Admin - a repo built on top of Material UI - as a possible solution. Word of mouth references are important and an upcoming live stream with Nico is planned. We appreciate the time of the participants and look forward to our next conversation.