The olympiad chemist teaching machines to make the hard molecules
Zlatko Jončev spent a decade earning the intuition of a great synthetic chemist. Now, as co-founder and CEO of b-12, he is trying to hand that intuition to a fleet of AI agents - and shrink drug discovery from years to months.
The best synthetic chemists carry something you cannot download. It is a feel for how atoms want to connect, which routes will fight you, and where a molecule hides its weakness. Zlatko Jončev spent the better part of a decade building that feel, one hard reaction at a time. Then he set out to give it away - not to another person, but to software.
That is the quiet ambition inside b-12, the company Jončev co-founded and now runs as CEO. It sits in Y Combinator's Summer 2025 batch, has three people on the payroll, and describes itself in plain terms: a chemistry copilot for pharma and biotech. Underneath the plain description is a large claim - that AI agents can plan and execute chemical synthesis well enough to compress early drug discovery from years into months.
Jončev likes to put the thesis in a single line.
The future of synthetic chemistry is agentic, and we're building it.Zlatko Jončev
01 - The problemWhy a molecule can take ten years
To understand why anyone would try this, start with the numbers that keep pharma executives up at night. A new drug can take more than a decade and north of two billion dollars to reach patients. Of every ten thousand compounds that enter the pipeline, roughly two make it out. Most of that attrition happens early, in the lab, where chemists design a route to a target molecule, test it, watch it fail, and try again.
The tooling meant to speed this up has its own problem. Expensive robotic labs sit half-used because the software to drive them is fragmented and hard to learn. A chemist who wants a robot to run a screen often has to become a part-time programmer first. b-12's bet is that the interface should be language, not code. Describe what you want to make, and the agents handle the route, the conditions, and the protocol.
Lab robots are everywhere. The reason they idle is that telling them what to do is harder than doing it yourself.
02 - The pathFrom olympiad medals to the bench
Jončev did not arrive at software through software. He arrived through chemistry, and early. As a student he placed second three years running at his national chemistry competition, then went three-for-three as a national champion and competed at the International Chemistry Olympiad, the sport's world stage for teenagers who can balance equations in their sleep. Serbia sent him to represent it.
The medals turned into a career. He earned a PhD in organic chemistry at the University of Basel, working on the kind of methodology - atroposelective synthesis among it - that rewards patience and a strong stomach for failed reactions. He trained as a medicinal chemist at Roche, one of the industry's most demanding schools, and spent time close to lab robotics companies where the machines he would later automate actually live.
The turn toward AI came at EPFL, where he did his postdoc in the Lab of Artificial Chemical Intelligence. There he worked on teaching large language models to reason about chemistry - not to chat about it, but to plan syntheses and pick apart reaction mechanisms with something closer to a chemist's strategy. It was also where he met Andres Bran, then a PhD student in the same lab, who had built one of the first AI agents for autonomous molecular planning.
03 - The partnershipTwo chemists, one idle robot
The founding story is unglamorous in the best way. Two people in the same lab, one doing a PhD and one doing a postdoc, kept circling the same frustration: the intelligence to plan chemistry and the machines to run it existed, but almost nothing connected them. They left academia together to build the connective tissue and named it b-12 - a wink at vitamin B12, one of the most notoriously difficult molecules to synthesize.
They named the company after the molecule that taught generations of chemists what "hard" really means.
The product line carries the same in-joke, borrowing names straight from the periodic table. Each element stands for one stubborn step in getting from an idea to a real compound.
04 - The proofOne plate, ninety-nine percent
Claims about AI in the lab are cheap. What Jončev points to is a pilot result that would make any bench chemist look twice: a screen that hit 99% yield on the first optimization plate, and an experiment that reached full conversion on a single attempt. In a normal lab that is weeks of iteration compressed into one shot. The company frames its route design as roughly 70% faster than the traditional back-and-forth.
Whether those numbers hold across the messy reality of pharma chemistry is exactly the question a young company has to answer. But the direction is clear, and it is the same direction Jončev has been walking his whole career - toward making the hard molecules, the ones other teams route around or abandon, because that is where the value quietly sits.
05 - The ideaChemical super intelligence
Ask Jončev where this ends and he reaches for a big phrase: chemical super intelligence. He means autonomous systems that can plan and execute complex, multistep syntheses with a human barely in the loop - not a smarter search box, but a colleague that happens to be software. It is an audacious framing from a three-person company, and he seems comfortable with the audacity.
There is a neat symmetry to the whole enterprise. A kid who won chemistry contests by out-thinking everyone in the room grew up to build a machine that could, one day, out-think him too. Most people would find that threatening. Jončev seems to find it interesting - the natural next move for a skill he cared enough about to want it to scale past a single pair of hands.
He trained for years to think like a great chemist. Now he is teaching the machine to do it, and betting the trade is worth making.
For now the work is early and the team is small, three people in San Francisco with a periodic table's worth of ambition and a name borrowed from the hardest molecule in the book. If it works, the chemists who come after Jončev may never learn the failure their way into intuition. They will just describe what they want to make, and watch the agents figure out the rest.