Sign In

David Vargas

Founding Engineer at Vellum

David Vargas is a Full-Stack Founding Engineer at Vellum AI, a company that specializes in LLM orchestration and AI development platforms.12 He joined Vellum as a founding engineer in July 2023, marking a significant career move for him.4

Background and Experience

Prior to joining Vellum, David had a diverse background in software engineering:

  • He started his career in 2014 at MIT Media Lab as a Software Developer.6
  • He has worked on various projects, including @samepagenetwork and @mark43.3
  • David has a strong foundation in software development, with 77 repositories available on his GitHub profile.5

Education

David Vargas received his education from the Massachusetts Institute of Technology (MIT)2, which likely contributed to his strong technical background.

Current Role and Contributions

At Vellum AI, David is part of a team working on cutting-edge AI technologies:

  • He is involved in helping companies deploy AI into production.2
  • David and his colleagues are exploring innovative ways to integrate AI tools like Cursor into their development process.1
  • His work focuses on making AI more accessible and practical for various teams within organizations.1

Online Presence

David maintains an active online presence:

  • He is based in Miami Beach, Florida.3
  • His personal website is davidvargas.me.3
  • He is active on social media platforms like Twitter (X) and LinkedIn.32

David Vargas appears to be a skilled and experienced software engineer who is now at the forefront of AI integration in software development, contributing to Vellum AI's mission to make AI more accessible and productive for businesses.

Highlights

Mar 28 · linkedin

You walk into a McDonald’s

You walk up to the cashier to place your order

Before you say anything, you see the manager start whispering to the cashier. You sit there waiting for a whole 5 minutes.

Finally, the cashier turns to you.

“What would you like?”

“Uhh, can I get a Big Mac?”

You wait another 5-ish minutes while the cashier just stands there thinking about what you just said, staring blankly at you. Why are they not doing anything? Finally, they allow you to pay. You have your waiting number.

As you wait, you see the manager is still there behind the counter. They whisper the same 5 minute monologue to each cashier before they take each of their respective customer's order. Shocked by this user experience, you down your burger, realizing that they actually gave you a Bacon cheeseburger instead of the Big Mac, and never return.

This is how I see a lot of system prompts designed today. It’s the easiest to define in software. Just stuff prompts with all of the possible context, and cross your fingers waiting for the magic to happen within the model underneath, only to still lead to suboptimal, bacon cheeseburger-esque results.

Instead of designing our prompts as if they’re just computers, what if we organized them as if they were employees?

  • How much context does the model really need before immediately serving the user?
  • How much of the menu needs to be loaded in order to know what to charge them?
  • Does the employee need to know the menu at all, or just how to use their menu-retrieving tool?
  • Do they need the full instruction manual of the menu-retrieving tool, or just know how to navigate it to answer their question?

The best classes in school teach us how to think, not what to think. I believe we should be designing our prompts in the same way.

Jan 23 · linkedin

How can we make LLMs more actionable?

The majority of the hype around AI today revolve around the virality of generative prompts. Prompts that are creating something fun and new based on uncommon groups of ideas. Yet for some reason when I log on to work, I still have to manually navigate through 8 different apps over the course of 90+ minutes just to get through my morning to-do list.

There are two very underutilized paradigms supported by some of the more popular proprietary models that should start to address this: JSON Mode and Function Calling.

The set of tasks we do at work are very well defined. All of the apps that we are performing these tasks on are backed by APIs that then manipulate the data that underly our clicks and keystrokes. Today, the interface that converts our user intent into a JSON that the application could consume is a UI consisting of a series of arbitrarily organized menus, tabs, buttons, inputs, and dropdowns. With JSON Mode and Function Calling, the interface could simplify to just a LLM-backed text input.

The question of when to use which comes down to architecture and stage of development. JSON Mode, currently only supported by OpenAI’s newest models, is best during early prototyping. It takes whatever directions given by your system prompt and the user input, and returns a JSON structured response based on what it thinks the schema should be. This is great for creating general knowledge bases or systems where you want the schema to be very flexible. But almost as hard to work with for outputs intended to be used by your application’s backend.

Function calling, also known as tool calling, is available in OpenAI’s, Google’s, and Anthropic’s latest models. They give the application developer agency over what schema the function call should adhere to. This crosses the gap directly from user’s intent to a backend API’s expectations. This approach is great for task specific and well-defined use cases.

Both of these approaches start to bring LLMs out of the Discord meme channels and into enterprise applications. They could help vastly simplify user interfaces by not requiring users to learn the application’s UI and tutorials in order to extract value. Instead, users could just act.

Related Questions

What projects has David Vargas worked on at Vellum?
How did David Vargas transition from MIT to founding Vellum?
What technologies does David Vargas specialize in?
How has David Vargas' role evolved since joining Vellum?
What are some notable achievements of David Vargas in his career?
David Vargas
David Vargas, photo 1
David Vargas, photo 2
Add to my network

Location

Miami, Florida, United States