QDC.ai: Shaping the Future of Optimization

From finding optimal routes in logistics to figuring out optimal schedules for public transportation or hospitals, we need to solve hard optimization problems for the world to run smoothly. 

However, solving these problems is really difficult. Even the existing optimization software is difficult to use and typically requires specialist knowledge (for instance, maths or physics PhDs) just to formulate the optimization problem. But what if there were an AI that would take the problem description from a human, turn it into equations, and then pick the right solver and backend to solve it? 

Founded in 2022 by Dominik Andrzejczuk, QDC.ai is lowering the floor on using optimization software and making a wide variety of optimization algorithms available, even applying methods from quantum computing.

Learn more about the future of optimization from our interview with the founder and CEO, Dominik Andrzejczuk:

Why Did You Start QDC.ai?

I spent the last eight years of my career in venture capital, which is how I came into contact with quantum computing. When I started at Morado Venture Partners back in 2014, we invested in Rigetti Computing. At the time, quantum computing seemed like pure sci-fi as there was no ecosystem, so Rigetti had to reinvent the wheel. 

Catapulted into this nascent industry, I absorbed all the new knowledge like a sponge and eventually co-founded Atmos Ventures in 2018. From the West Coast of the United States to the Far East of Europe, we looked into a lot of quantum computing companies and eventually invested in ORCA Computing and Oxford Ionics

While everyone in Silicon Valley is laser-focused on product and user experience, I learned that most quantum software companies were only focused on delivering R&D projects rather than real products, especially when selling to large corporations or banks. As William Hurley of Strangeworks once put it: Even if a ten-million-qubit quantum computer appeared in New York City today, no one would know how to use it. This stuck with me and got me to ask myself: What areas will be first affected by quantum computing? Certainly high-performance computing, and in particular solving optimization problems!

Today, optimization problems are not only hard to solve, but the existing optimization tools are also hard to use. Even the most sophisticated machine learning teams do not use optimization tools like Gurobi or Google OR-Tools as they require just too much effort. They always need consultants—math and physics PhDs—to do the mathematical modeling of the optimization problem. Thus, a large part of the market is underserved, and many processes in the world could be better optimized. 

Optimization seemed both an important enough topic and one of the first application areas of quantum computers, so I decided to start QDC.ai to make the world run more optimally. 

How Does Optimization Work?

Our goal with QDC.ai is to build something customers immediately get value from and the developer experience is spot on. Like Keras for deep learning. A product that is sticky and developers can’t live without. A pain killer, not a vitamin. 

After speaking to lots of developers, we figured that we needed to develop an AI that translates a natural language description by the developer into partial differential equations describing the optimization problem. Like ChatGPT is an interface to use GPT, we’ll build a language interface to solve optimization problems. By focusing on a narrow market segment in logistics and supply chain, we are going to build such an AI for a few use cases and combine the semantic description of the optimization problem with real-world data from the data pipelines of our customers. 

Currently, no package encompasses all the relevant optimization algorithms. We don’t want to reinvent the wheel, so we’ll offer plugins for Gurobi and Google OR and eventually provide about 20 different types of optimization algorithms. Our AI will be intelligent enough to pick the right solver for a given problem, whether it’s a mixed integer or QUBO problem.

Machine learning will also come in handy during the optimization process to improve customers’ models. While raw machine learning can work with random data to figure out the function to model that data and predict the future, it needs a ton of data to make useful predictions—its strength is also its weakness. So it’s better to start with a model based on first-principles maths or physics: Like traffic in a city can be modeled as fluid flow, although it does not behave exactly like a fluid, machine learning can be used to map the discrepancies between such a first-principles model and reality to help find a much more optimal solution.

Beyond making optimization software easier to use through machine learning, we also need better backends to solve optimization problems, as these tend to be very computationally expensive. And that’s why we’ll look into quantum computing, which will likely have its first impact in the realm of optimization.  

Like TensorFlow started out only using CPUs and then released plugins for GPUs and other hardware-accelerated chips like Google’s TPU, we are currently focusing on providing business value and a great developer experience based on existing chips. Additional backends—like a quantum one—could be later imported, just like a library. Ideally, the backend gets chosen automatically, without the user even noticing and without needing to rewrite code. 

That means we’re not dependent on quantum computers to be around to provide business value. But to stay hyper-competitive even in the long run, we’re thinking about employing them early on. We’ll determine the specifics, for example, which quantum algorithms we’ll use, through further pilot projects and based on customer needs. Our R&D around quantum computing will be very much tied to real business applications. 

In 2022, ORCA Computing demonstrated that their photonic, ‘boson sampling’ quantum computer could solve a QUBO problem better than a D-Wave advantage device and a classical benchmark. This is the first evidence that quantum-inspired algorithms might actually be better than classic optimization algorithms, that is, more performant, less error-prone, and give better solutions. 

Besides quantum annealing, we are also looking into quantum reservoir computing, where you use the error proneness of quantum computers to your advantage to train machine learning models with fewer data; there’ll be lots of applications even for noisy quantum computers! 

How Did You Evaluate Your Startup Idea?

Many founders try to go for venture capital (VC) funding just based on an idea, but that’s a mistake. Venture capitalists want to see that an idea can have clear business value. They’re not into funding science experiments. They like to see initial traction from the market.

When we started our journey in April 2022, our goal was to close a first customer before looking for VC funding. We spoke to tons of developers, AI engineers, and physicists at logistic companies before we even started building the product and IP, not to waste time on technical details. 

We figured that traditional logistics companies have lots but messy data and that we would have to first build plumbing for their data pipelines before we could do any useful analysis. So we decided to ditch the traditional logistic companies for now and try to find customers with clean data that we could use directly and focus on optimizing their processes.

On a business trip to California, I met a few tech unicorns working with airlines and large logistics companies. We learned that by working with these tech companies instead of traditional logistics companies, we could start working with clean data and provide business value within a few months. 

That’s why it’s so important to talk to lots of potential customers first. Our initial thesis that we should work with traditional logistics companies turned out not to be the best approach. So we adapted it, helping us construct a more coherent story. Only once you have a story that makes sense should you talk to VCs. 

What Advice Would You Give Fellow Deep Tech Founders? 

First, make sure you can pitch to a complete industry outsider. Your knowledge about the problem, the tech, and the industry will be way above that of the average Joe, but you still need to be able to break them down so that the average Joe can understand. There are a few highly sophisticated deep tech investors that will understand and many deep tech investors that will just pretend to have understood what you said. 

Second, focus on how your solution will create value. Most techies are deep into the details of their technology. But in the end, someone needs to buy a product that’s actually making someone’s life easier.

Last but not least, many founders coming from academia are not very good at branding and might consider marketing even below their line. Yet, the truth is that great tech doesn’t sell itself, and it’s valuable to think outside the box on how to promote your company. For example, for QDC, I hired a TikToker as marketing director—not because he’s knowledgeable about quantum or scientific computing, but because he knows how to grab people’s attention. And that’s incredibly valuable as you figure out how to sell your product and minimize your costs of sales. 

Many quantum companies pay PR companies lots of money for PR articles, but few people actually read them. When you ask scientists, they would tell you that the claims of most of these PR articles are scientifically dishonest or at least inaccurate. 

Even if you’re in a highly sophisticated market segment like quantum computing, your customers will still be humans. They need to laugh as well! They, too, have sons and daughters. And they, too, want their sons and daughters to think what they’re working on is cool. So, you have to brand yourself as something cool. 

Comments are closed.