Bagel, Monetizable Open Source AI 🥯

CV Deep Dive

Today, we’re talking with Bidhan Roy, Founder and CEO of Bagel.

Bagel is an AI research lab leveraging advanced cryptography to make open-source AI monetizable. Their platform, The Bakery, empowers contributors to collaboratively build AI models without exposing model parameters or data to one another. It ensures proper attribution and fair revenue distribution to all contributors. Bagel's core technologies include privacy-preserving machine learning and a novel parameter-efficient fine-tuning (PEFT) technique, among others.

Bagel serves early-stage AI startups and researchers who rely on open-source AI but need superior resources like datasets, base models, etc. By enabling secure collaboration and monetization opportunities, Bagel addresses a critical challenge for open-source AI developers.

In this conversation, Bidhan explains how Bagel came to be, the technical hurdles involved, and what the company plans to focus on next.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Bidhan 💬

Bidhan - welcome to Cerebral Valley! First off, give us a bit about your background and what led you to found Bagel? 

Hey there! My name is Bidhan, and I’m the founder of Bagel. My background is in engineering, and I’ve been working in machine learning for over a decade now. Previously, I led a team at Amazon Alexa that was responsible for the machine learning infrastructure supporting over 100 million devices worldwide. So, we’re talking about one of the largest machine learning infrastructures in the world. I've worked in a few other places, like Cash App, Instacart, etc., and throughout that time, I’ve also been a big contributor to open-source projects.

I love open source AI, everything about it. But the question that's always been on my mind is, how do we make open source monetizable? Open source is essentially charity right now, and that’s not sustainable, especially as we move into this new AI age, where open-source AI plays such a crucial role in the rapid development of the field.

That’s a problem I’ve thought about for a long time—how do we make contributing to open source worthwhile and incentivize others to get involved? Then, back in 2017, when I was working at Cash App on their Bitcoin trading app, I was introduced to Bitcoin and, through that, to a lot of other cryptography concepts. That’s when I had my “aha” moment: we can use cryptography to monetize open-source AI. That’s where the idea for Bagel started to form.

At Bagel, if I had to sum it up in one line, we’re making open-source AI monetizable using cryptography. That’s the core of everything we’re building. I started Bagel about a couple of years ago, right when open-source AI really started to take off. It was the pre-ChatGPT era, and it was clear to me that open-source AI was going to be central to AI development. I wanted to embark on the journey to make it monetizable.

Give us a top level overview of Bagel - how would you describe the startup to those who are maybe less familiar with you? 

We envision Bagel as a foundational breakthrough that makes open-source AI monetizable through advanced cryptography. To achieve this, we concentrate on three key pillars:

First, we solve for attribution. Every open-source contributor working within the ecosystem gets properly credited for their contributions, and there’s a system in place to distribute revenue to them.

Second, we tackle privacy. Right now, open-source models and datasets contributed to popular platforms are completely open, so even if contributors want to monetize, they can’t. We allow contributors to collaborate in machine learning without revealing the privacy of their resources. This concept is part of a research area called privacy-preserving machine learning, or PPML, which we leverage in the Bagel ecosystem.

Third, we focus on performance. Many open-source contributors don’t have access to high-quality resources like GPUs, unlike the top AI labs. We use cryptography to enable resource owners, like data centers around the world, to provide their resources to open-source contributors. This allows them to fine-tune their models and create better versions using those resources.

So, to summarize, we’re solving the monetization of open-source AI through three pillars: attribution, privacy, and performance. That's what we do.

Talk to us about your users today. Who’s finding value in what you’re building with Bagel? 

Bagel today mainly serves early-stage AI startups. These are the ones that rely heavily on open source because they don’t have all the resources that bigger companies like OpenAI do. OpenAI doesn't need us because they already have access to everything they need. But these smaller AI startups really depend on high-quality base models, datasets, and compute power to build what they’re working on. 

Plus, they're the ones who need to monetize the most—they can’t afford to just give everything away for free. That’s where we come in. Our target audience is primarily early-stage AI startups, but we also focus on early-stage researchers from universities and even hobbyist developers who are building on open source.

As far as cryptography and its application to AI, which use-cases do you see as the most powerful? 

I believe cryptography acts as a collaboration tool, allowing people to work together without ever meeting in person or building trust through traditional means. Let me give you some context: we recently launched our first product, the Bakery. The name comes from the idea that it's where bagels are built, and it’s a platform where developers can fine-tune their models using resources owned by others—like datasets, base models, and even GPU power. 

In the Bakery, you might have one person with a proprietary dataset, another with a base model, and someone else with GPUs located in South America. None of these resources belong to us; they belong to the individual contributors. What cryptography does is allow all these parties to collaborate without revealing their sensitive information. I don’t have to reveal the weights of my model, you don’t have to expose the private details of your dataset, and the person providing the GPUs doesn’t need to disclose anything either. 

Cryptography makes it possible for geographically distributed people to work together while keeping all their resources private.

There’s been an explosion of interest in agentic workflows and multi-modal AI. How has that shaped the way you’re thinking about building Bagel? 

Bagel’s solution is generalized across a wide range of AI startups, but we do have teams that are building agents. One example is a team called Talky AI. They’re creating a conversational agent for enterprises. Essentially, you can call a number, and the agent will answer questions about whatever enterprise they’re working with. They use Bagel for model fine-tuning, gaining access to high-quality GPU compute and proprietary datasets through our platform. So while the base infrastructure supports a variety of applications, agent-based solutions happen to be a big part of it because they’re just popular right now.

What’s been the hardest technical challenge you’ve had to overcome to get Bagel to where it is today? 

One of the biggest technical challenges, which also happens to be a core part of our business model, is the privacy aspect. How do you let geographically distributed people collaborate without giving up their proprietary information—whether it's a dataset or model parameters? That was one of the hardest problems we had to solve, and we’ve managed to figure it out.

In short, Bagel uses two key technologies for privacy-preserving machine learning. The first is called fully homomorphic encryption. Essentially, it allows you to perform computations on encrypted data and get the same results as if you were working with unencrypted data. It sounds like sci-fi, and it’s definitely cutting-edge, but this is the future of privacy.

The second is differential privacy. It works by adding a controlled level of noise to the data, which helps keep it private. There are pros and cons depending on how it’s implemented, and while we haven’t fully published our method yet, we allow model fine-tuners to add differential privacy to the stochastic gradient descents they generate. This method, known as DP-SGD, ensures better data privacy during model fine-tuning. We’ll be releasing more details about this soon.

How do you plan on Bagel progressing over the next 6-12 months? Anything specific on your roadmap that new or existing customers should be excited for? 

Our focus is always on two key areas: product and research, both of which are critical in AI. With the recent launch of the Bakery, we’re growing rapidly. The next six months are all about meeting the huge demand we’re seeing. The combination of our privacy features, cryptographic collaboration capabilities, and monetization approach has sparked an incredible amount of interest. So, our immediate priority is to keep up with that demand.

What users can expect from us is the ability to fine-tune their models with just one click on the Bakery platform. We’ll be opening it up to the public very soon.

On the research side, we're committed to open source and collaboration. We’ll be open-sourcing our cutting-edge research on how to use cryptography to monetize open source AI and how to enable collaboration without compromising privacy. Users and researchers can keep an eye on our research blog and social channels to stay up to date with everything we’re doing.

Lastly, tell us a little bit about the team and culture at Bagel. How big is the company now, and what do you look for in prospective team members that are joining?

We're a team of ten people, based in New York, though the team is distributed. Collectively, we’ve designed and run some of the biggest machine learning systems in the world. Together, we’ve published more than 30 original papers on machine learning, cryptography, quantum computing, and more. We also have team members with experience in venture capital, having deployed over $20 million in various cryptography startups.

And yes, we’re hiring! We’re particularly looking for engineers and researchers interested in the intersection of AI and cryptography. If that sounds like you, we’d love to chat. Please reach out to [email protected].

Conclusion

To stay up to date on the latest with Bagel, learn more about them here.

Read our past few Deep Dives below:

• Our chat with Datastax's Chief Product Officer, Ed Anuff ❇️

• TensorStax is building Autonomous AI for MLOps 🌐

• Dashworks, your AI-powered internal search tool 🔍

• Delphi - the verified digital cloning platform 👥

• CodeGPT is intertwining human language with code 🌏

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.

Join Slack | All Events | Jobs