mdw.la :: Matt Welsh

Projects

Published: at 07:58 PM

Here are some of the non-research projects I’ve been involved with. See my research page for more information on academic work.

Large Language Models as a new computing platform

My current fascination is with the potential of Large Language Models (LLMs) as a new computing platform. LLMs like GPT-4 are showing the signs of being able to solve problems and complete complex tasks using a natural language interface, and I believe that this development points to a new kind of natural-language computer.

My first startup, Fixie.ai, was originally working on a developer platform to help companies leverage these powers of LLMs in their own applications. We raised $17M in seed funding from top-tier VCs in early 2023, and had thousands of developers building on our platform. I left Fixie after the company pivoted into a very different space in early 2024, but I am continuing to work in this space.

Podverse - AI superpowers for podcasters

I built Podverse, a site that automates the process of generating transcripts, summaries, and AI chatbots from podcasts.

I listen to a ton of podcasts, but it’s frustrating to find relevant episodes, or even content that I had listened to previously. I also know that podcasters want to take advantage of the latest AI tech to connect with and expand their audience, but they’re not usually equipped to bring all the pieces together. I wanted to build a complete, end-to-end solution for podcasters that is easy to use.

All you have to do is provide your RSS feed URL, and Podverse does the rest — automatically ingesting new episodes as you release them, generating transcripts and summaries, identifying speakers, and populating a custom chatbot that knows about all of the content of your podcast.

Optimizing ML models for deployment

I worked at both OctoML and Xnor.ai on optimizing and deploying machine learning models. OctoML was founded by the creators of Apache TVM, an optimizing compiler for ML models. TVM takes a deep learning model and compiles it into highly efficient and auto-tuned machine code for the target hardware. Apart from TVM, at OctoML I was involved in building their cloud-based service to measure, tune, and package ML models for a wide range of hardware backends.

At Xnor.ai, I worked on optimizing ML models (primarily for machine vision) to run efficiently on embedded devices like microcontrollers. When Xnor was acquired by Apple, our team’s technology was integrated into Apple’s on-device AI platform.

Bringing the web to the next billion users

I spent 8 years at Google as a Principal Engineer and engineering director for the Chrome Mobile team in Seattle. Our team focused on making Chrome and the web great for the next billion users in markets such as India, Africa, and Southeast Asia. We developed new browser capabilities and cloud services to make the web faster and use less data.

Some of my team’s projects at Google included:

Personal projects

I love hacking and building stuff. Check out Team Sidney Enterprises for some of my personal projects.

Here is my GitHub profile.