This week I caught up with Edvard Sivickij. He was a design engineer in aviation. Then he studied data science at Turing College, became a Junior Team Lead with us, worked as a data scientist at Kilo Health and an AI engineer at Nexos AI, came back for our AI engineering programme and founded Condense.
Condense sits between AI agents and the models they call, and cuts how many tokens each request uses. Edvard has just been invited to the a16z speedrun founders bootcamp. Out of about 4,500 applicants, roughly 20 got in.
Starting with drones
Edvard has had a side project since his last year of school. As a design engineer he worked on different types of UAVs, including morphing wings. Around 2018 to 2020, the only thing companies were allowed to do with drones was film events.
He tried to start his own drone company several times. The answer was always that it was too early. His team only looked at Europe, because that's where the money was. At the same time Zipline was flying drones in Africa. There was no market there, but there were also no standards. "You just launched," he said. "If a drone fell somewhere, nothing happened."
He also saw that even a drone company runs on software. So he started learning to code.
Learning at Turing College
He taught himself the basics first, then joined Turing College to study Python and data science. Afterwards he did a master's in ML and AI. He calls our programme the practical side and the master's the theory.
He was one of our most active learners and later became a Junior Team Lead. He took part in the hackathons we organised and in plenty we didn't. We introduced him to Kilo Health, where he started as a data analyst and became a data scientist. Then he joined Nexos AI as an AI engineer.
I asked which programme helped most with getting hired. He said data science, because Nexos AI was looking for ML engineers with strong development skills. "Data science is the core of everything," he said.
For founding a company it was different. Edvard says the AI engineering programme was the last missing piece he needed to succeed as a founder and to find a niche that works.
I agree. Skills build in layers, so the fundamentals matter most. The other part is what people do outside formal learning. Edvard puts it this way: the more you learn and take part, the better your chance of being in the right place at the right time.
Taking the long way
While at Kilo Health he started building his own agents. The first ones ran on GPT-3.5. They looked amazing and did almost nothing, because the models understood too little.
It took him longer than most to start a company. Part of it was living in Italy, where he says the startup scene was little more than Bending Spoons. In Vilnius you can find people who have made the mistakes before you. In Italy he had to make them himself.
One of them was a learning app. His team spent a year building it and then presented it to us. The feedback was harsh. He says the lesson was to show something to people much earlier.
When he joined our AI engineering programme last summer, he was already using Cursor. A developer now on the Condense team, who had always written his code by hand, called Cursor nonsense and said it would never replace people. Today that developer runs five or six agents in parallel. People need proof that something works. Once they have it, things change very fast.
What Condense does
Condense started out building its own models for specific tasks. The team dropped this because models are easy to switch. "Yesterday you had Opus, today you plug in GPT and things keep working," Edvard said. The harness around a model is much harder to leave, and that is where everyone started building.
Condense now runs as a proxy between an agent and the LLM. It manages cost and caching, compresses context so each request carries fewer tokens, and tracks how conversations develop. The next step is to use that data to give teams insights and help them improve their agents.
Edvard calls himself a generalist. He says things got better once he found someone for the team with deep expertise in one niche. In his view a general-purpose agent is easy to replace with OpenAI, so an AI startup needs a specific problem. Condense chose developer tools, because the team knows those problems from its own work.
The team keeps the consumer product as a lead magnet. Individual users don't feel much pain from token costs, and each one has to be convinced separately. The team has worked on Condense full time since the start of summer. Condense has raised a six-figure round, with investors from Lovable.
Hiring and building from Europe
Condense uses fractional hires for now. The first full-time roles will be in go-to-market and marketing, where the team is weakest. Edvard wants to hire a strong salesperson and learn by working next to them. "Sales has to be founder-led," he said. "You can't hand it to someone and say go for it."
For engineers, he cares more about how much someone wants to build than about years of experience. He wants the team to stay small. I hear the same from founders of larger companies. Most of them wish they could go back to a team of under 100 people, where everyone knows each other and the founder is still hands-on.
He thinks NordVPN and Vinted show that big companies can be built from Europe. He points to Stockholm, where founders work weekends. The hard part is regulation. Data privacy rules make training models difficult.
In my view it depends on who the customer is. A consumer product can be built from anywhere. B2B usually needs a hybrid model, with part of the team where the customers are.
His advice
I asked what he would do if he were back where he was seven years ago. "You can't predict what will happen," he said. "You can only decide based on your current circumstances." Look at what you can do now and where you want to be, then decide.
He would learn to code again. Building things himself was what mattered to him, and when he didn't know how to do something, he learned it.
He puts most of his results down to stubbornness. For some founders it takes longer than for others. It reminds me of Akio Morita, the co-founder of Sony. In his book Made in Japan he wrote: "The challenge is great; success depends only on the strength of our will."
Try Condense
Condense works as a drop-in proxy for Claude Code, Codex and OpenCode, and you keep your existing provider key. On one long coding session, Condense reports cutting the bill by 72% (TestingCatalog).
You can try it at condense.chat. If your company is spending thousands per month on AI credits, definitely try it.
Feel free to contact Edvard on LinkedIn.
Want to start building with AI?
Ship real products with AI agents on our hands-on Building with AI programme.