Reasonance: Shaping the Future of End-to-End Machine Learning Workflows
Over the past five years, machine learning has evolved from a curiosity to a transformative force in various industries. Its potential to enhance decision-making and automate processes is now widely recognized.
However, many companies face significant challenges in harnessing the power of machine learning. Whether developing their own models or leveraging external ones, organizations often grapple with data ownership, security, and compliance issues. These obstacles can impede their ability to capitalize on the benefits that machine learning promises.
Reasonance was founded by Todor Kostov, Konstantin Tsenkov, and Manuel Lang in 2020 to help customers implement enterprise-grade machine learning and analytics solutions, with over 30 successful projects completed and a growing list of industry-leading customers. In addition, they have built the MLOps platform ATLAS, helping organizations streamline their machine learning projects.
Learn more about the future of end-to-end machine learning workflows from our interview with the co-founder and CEO, Todor Kostov:
Why Did You Start Reasonance?
After graduating from Karlsruhe Institute of Technology (KIT) in computer science, I started working as an independent contractor, looking to do something cool with technology. I’ve been into machine learning and statistical analysis since 2012, doing various projects in the chemical, energy, and transportation industries.
From the first few projects, I recognized that 90% of a machine learning project’s workload is always the same: data integration and storage, API creation, information routing, and all the data plumbing no one likes to do. Only 10% of the work is playing with the weights of a machine learning model. And most companies lacked the capability and structure to deal with the boring stuff. Maybe there was an opportunity to develop a product!
My co-founders and I started developing ATLAS, a platform for end-to-end machine learning, to help companies move from Jupyter notebooks to having solid infrastructure for their machine learning projects and reduce the workload as much as possible. This idea developed in 2019, and in 2020, we founded Reasonance to put it into practice.
At the time, not many people were using machine learning. Many pilot projects were ongoing, but few projects made it into production. This has completely changed over the last four years. Machine learning went from being a curiosity to seeing mass-scale adoption, with lots of companies suddenly needing support to deploy and maintain all these machine learning models. While cloud infrastructure has become extremely flexible, most companies need help from consultants to maintain their cloud and machine learning infrastructure; that’s why bootstrapping has worked so well for us, being profitable from day zero and growing through customer revenue.
How Do You Help Companies With End-to-End Machine Learning?
Our platform ATLAS is a plug-and-play solution for all of a company’s problems around data ingestion, storage, and processing for building machine learning models. Data comes in, and we’ll take care of everything from storage, data querying through SQL, and data processing pipelines all the way to receiving results. Using ATLAS, organizations can reduce the cost and time-to-market of their machine learning project and ensure that their ML solutions make it into production.
For many companies, an on-premise solution is a must. But many don’t have the financial resources to support this. Think of someone like the public utilities in Germany who would like to set up customer support automation as a chatbot. It would take them far too much effort to manage compliance, hire legal people, and read through the terms and conditions if they wanted to use ChatGPT. Instead, they expect everything to work on-premise so that sensitive data never leaves their system.
Often, running on-premise is also cheaper. We helped one company specializing in big data, which had a forecaster set up as a cloud solution running 24/7, leading to operational costs of around 20K/month. We migrated it to on-premise, switched the tech stack, set up a simple Kubernetes architecture, and brought costs below 1K/month.
Most machine learning setups don’t have to be fancy, exotic, or expensive. It depends on your requirements: stream processing of high-velocity data differs completely from batch-processing big data.
Our end goal is to minimize the consulting we need to do to zero and make ourselves obsolete. Consulting scales only with the number of people. Finding good people is hard, and every person can only spend so much time. Initially, it is crucial for any startup to learn from its customers, and consulting plays an important role in this.
Our offering changes with the size of our customers. For small companies, it’s mostly consulting and managed turn-key solutions, for mid-sized ones, it’s mostly consulting and the ATLAS platform, and for large companies, it’s mostly our software platform. Big players have a lot of internal know-how, and they don’t want consultants to learn about their IP, as it’s their competitive advantage.
How Did You Evaluate Your Startup Idea?
We came from machine learning and data engineering and worked with different companies in media, energy, and other fields. As a consultant, you see all of their problems. At some point, you recognize specific patterns and think about a product to address a subset of them. For this, it really helps to be boots on the ground, learn about your customer’s pain points, and see what solves the problem.
For example, when you talk to manufacturing businesses, they have steady processes established already, and you get to work with and around them. It’s very much about finding out what a particular step costs in a process, finding a better and cheaper way to solve it, and providing measurable results that you’re better—and then selling that improvement to more customers with a similar process.
It’s often not just a financial problem of saving costs but really understanding where you integrate with an existing process. For example, predictive maintenance makes no sense in many cases, and if it’s worth it depends on how much time and money you’ll save.
You get the best machine learning models from domain knowledge. For example, in manufacturing companies, the managers and production engineers have immense domain knowledge, which beats data volume every single time: if you can get models to incorporate domain knowledge, you reduce the data requirements as well as the model complexity, which in turn makes the solution more interpretable and less of a black box.
What Advice Would You Give Fellow Deep Tech Founders?
Coming straight out of academia, I had very little business experience. I had never been in a room where decisions were made. But I learned everything from sitting with executives in meetings—here is where consulting becomes very valuable.
Build your network and find a proper way to communicate with customers. Figure out the right amount of technical and business details to share. Structure your offering so your customer can understand it; don’t overpromise, but also don’t undersell—promise what you can confidently deliver. That is hard to get right.
Many venture-backed startups go to market with a product that is not yet ready. But after bootstrapping for a couple of years, we have developed a product that we can trust, knowing what we can do and at what price.