Leeroo: Trainable AI Agents That Learn Like Humans

From NLP Research to Startup Founding

I originally left my home country at around the age of 22 and moved to Switzerland to pursue a PhD at EPFL, where I focused primarily on natural language processing and AI. I also spent time at several other institutions in Switzerland, including the University of Zurich and IDIAP Research Institute. After four or five years in academia, I transitioned into industry, continuing to work in generative models. One of the more notable places I worked was Meta AI.

While at Meta, I met my co-founder Majid—we collaborated on a project that, at the time, wasn’t called an “LLM-as-a-judge,” but the concept was similar. We were using AI models to evaluate the output of generative models. That particular model has since gained a lot of traction—it’s been downloaded over a million times on Hugging Face.

I think at the moment, there isn’t a significant gap between science and production—at least in the LLM and AI domains. As soon as a paper comes out, it’s often just a matter of days before someone implements it in production. It wasn’t always like this. In the past, it followed the more traditional model—where you'd conduct research, and it might take years before it translated into production. That shift really influenced how I thought about my career.

By the end of my PhD, I was certain: rather than going back to Big Tech, I wanted to start my own company. And it just felt like a golden moment—there are so many tools now that enable you to build products end to end, and I already had the technical depth in LLMs and that broader domain. So I thought, this is the right time to take a bet on a startup. Eventually, Majid, Arshad and I decided to start Leeroo, and we brought on a third co-founder from India, who focuses more on low-level systems and engineering. That’s how we came together.

The Problem: Training AI Agents into Experts

For us, the core problem we're addressing is the disconnect between human domain expertise and what current LLMs or AI agents are capable of.

There are experts in every field—medicine, finance, software engineering—people with deep, valuable knowledge. And then there are large language models, which are generalists. These models are improving rapidly, and sometimes they can even produce expert-level output. But there's no effective way to transfer a domain expert’s knowledge into the model. It’s not predictable or systematic.

The Solution: Continuous Learning AI Agents

What we’re trying to do is bridge the gap between human expertise and agentic AI capabilities.

We want to empower experts to directly and continuously transfer their knowledge into agents—so the agent becomes more like an expert colleague. That’s why we call our agents continuous learning agents. You can interact with them daily, update them with new insights or corrections, and over time they improve—without needing to collect massive datasets, retrain, or fine-tune using heavy compute. Instead, we’ve built a module on the backend—what we call the learning engine—that observes interactions and updates the agent’s knowledge. It reviews everything that happened today and decides, “These are the things I should learn for tomorrow.” That’s our vision: creating AI agents that grow and adapt just like a teammate would.

We're starting with the domain of data and AI—so the first agents we’re building and deploying are for roles like data engineers, data analysts, and machine learning engineers. These agents are designed to operate at an expert level within those specific domains. Once we validate that, the plan is to expand into other areas.

For example, we’re currently working with a company focused on private banking. They shared with us a training package that they normally give to interns to help them prepare for their presentations during their time there. We took that same material and gave it to one of our agents—essentially training it to act like an intern in that specific context. It’s still experimental, but it shows the broader vision: allowing agents to quickly absorb domain-specific knowledge and become high-performing contributors within any field.

3 Ways We Validated the Idea

The idea for Leeroo has evolved quite a bit over time. Our latest direction—and the one we've built Leeroo around—is what we call continuous learning agents. The validation for this idea came from both our own research and from observing broader trends in the research community.

As researchers, we saw that others were also starting to explore the concept of self-learning agents. Some are using reinforcement learning (RL) algorithms, others are adding memory modules to their agents. Memory, in particular, is becoming a popular direction—some companies are even building products that add memory layers to tools like browsers. So there are multiple efforts converging on the same goal: enabling agents to learn continuously, become more self-aware, reason more effectively, and eventually reach expert-level performance in specific domains.

From a validation standpoint, we’re seeing momentum both in academia and in industry. Internally, we've also run experiments and benchmarks—particularly in areas like complex reasoning and question answering. We plan to publish our findings soon in a technical report, to contribute back to the research community and share our perspective on how learning agents can evolve.

So overall, the validation has come from three directions:

1. research trends,

2. product interest, and

3. our own experimental results.

Building the MVP for Enterprise Customers

In terms of how we built the MVP technically, we took a pretty direct approach. We didn’t rely much on multi-agent frameworks or any of the existing abstractions. Instead, we worked at a lower level—calling SDKs directly and building things from the ground up based on what we needed. That gave us more control and flexibility early on. In our product, the agent is connected to a learning engine which makes the agent perform better over time by leveraging its learning.

Of course, working with enterprises also introduced constraints—mainly around security and deployment. Many of our customers require the product to run fully on-prem, inside their VPCs, or they expect strict compliance certifications like SOC 2 Type I or II. That’s been one of the biggest hurdles—just meeting those requirements and navigating enterprise procurement processes. Nonetheless, we have designed the architecture of our product in such a way that it passes the security process in around 2 meetings.

Another challenge is access. Unlike public tools or open-source platforms where users can immediately connect to their data sources, enterprise environments are more locked down. We often have to work closely with internal teams just to get the necessary permissions, which slows things down a bit. But it’s part of the territory when building for the enterprise.

Iterating Based on a Customer Feedback Loop

Right now, we’re mainly selling our product to large enterprises, and that’s shaped a lot of how we approached the product. One thing that helped early on was the high signal we received from these customers—especially around the learning aspect of the agent.

Every time we added a new piece of knowledge to the agent, we would pass it along to the customer and get immediate feedback. That fast feedback loop allowed us to validate what the agent had learned and iterate quickly based on real-world usage.

Business Model and Pricing

Currently, our business model is more like a software-led growth approach. We sell our product and services as a complete package to large enterprises.

Right now, we’re selling agent seats. Once a company has a few agents deployed, we work on expanding within that organization by adding more agents. 

One of our ideas is to sell agents specifically to companies that are hiring for data engineers, data analysts, or machine learning engineers. The interesting part is that these companies can interview the agent just as they would a human candidate. If the agent passes the interview, the company can hire it.

This approach positions our agents as both cost-effective and efficient, which can complement the work that human hires do. It’s more about setting the agent at an expert level in that specific domain, and our contracts are based on performance. We guarantee that the agent maintains expert-level capability throughout. Essentially, we say: you interview the agent, and you decide if it meets your standards

Acquiring the First Few Customers

Our first paying customers came through connections we had with some C-level executives from our previous experiences and networks in these companies.

For one of them, we identified a significant problem they were facing around data cleaning, data quality, and their entire ETL (Extract, Transform, Load) pipeline. At the same time, they were looking to hire a full data engineering team.

We approached them and said, “Hey, you don’t need to hire the whole data team—our agents can be your team.” That’s how we started selling.

To date, we’ve sold them three agents—the data engineer, data analyst, and machine learning engineer—and they became some of our very first paying customers.

Marketing as a B2B Agentic AI Startup

Honestly, we’re more researchers than marketers, so marketing hasn’t been our strongest suit yet. That said, we’re starting to build up our efforts. Right now, we’re focusing mostly on LinkedIn since that’s where most enterprise decision-makers are active.

We’re also working on creating teaser demos of our agents, which we hope to release soon to generate more interest.

Managing the Co-Founder Relationship

One of the great things is that before starting Leeroo, we were already friends for about three or four years. That foundation helped a lot—it means we can talk openly and transparently with each other in a friendly environment. We feel comfortable discussing issues as well as successes without hesitation.

In terms of work, we’ve mostly divided responsibilities based on our strengths. I focus primarily on technology and research. Our co-founder Arshad handles lower-level infrastructure and coding, and Majid, our CEO, leads customer relations and fundraising.

This division helps us each stay in our zones of expertise while having clear interfaces to collaborate. Because of this, we’re able to build on each other’s work efficiently—for example, if I come up with a tech idea, the infrastructure team can quickly push it into production with just a few clicks.

This setup saves us a lot of time and creates a nonlinear impact—we don’t each need to be experts in every part of the codebase, yet we move fast together.

Weekly Workflow

We have a short daily stand-up every morning, usually around an hour. Tasks are generally divided based on our roles—if it’s research-related, I’ll handle it; if it’s infrastructure or customer-related, the relevant co-founder takes over.

That said, we’re also flexible. If a task comes up and someone has more capacity, they’ll jump in regardless of role. So on a day-to-day basis, it’s pretty fluid and collaborative.

At the end of each week, we do an internal product demo. Each of us presents what we’ve worked on, and we use that time to identify issues and areas for improvement. That session naturally leads into our weekly planning—it helps us align on what to focus on next.

What Keeps Us Going: Impact

Even today, the pressure and workload are intense. But every morning, I wake up excited—because I love building what we’re building.

What drives me is this belief that we can have a real, meaningful impact on humanity. Imagine a world where anyone can take their knowledge, turn it into an executable agent, and offload repetitive or mundane tasks. That would free people up to focus on higher-order goals—like creating, discovering, curing diseases. That’s what I mean by raising the upper bound of human potential.

No one should have to waste their life on repetitive work. People should have time to rest, enjoy life, and ideally, contribute to something bigger—like the researchers using AI to cure cancer. That kind of work inspires me deeply. It shows what humans are capable of when we’re not bogged down by busywork.

And on a personal note, this is why I turned down offers from big companies before joining Leeroo. I didn’t want to just do something safe—I wanted to do something with impact.