Company

One hard prediction.

Heuresia exists to predict whether a molecule will be toxic before anyone has made it. Everything here, the team, the data, the rules and the roadmap, is built to make that one call trustworthy.

Team

Mark brings the medicine and the faculty link; Filip brings the capital and the operations. Around them, a small team of engineers and scientists in four teams, every one converging on the same proof.

Mark Sobolewski
Founder & CEO
Started Heuresia as a medical student at the Medical University of Warsaw. Runs the company, keeps the engine honest, and brings the medicine and the faculty connections that turn a benchmark result into something a toxicologist will stand behind.
- Full bio
- LinkedIn
Filip Jojnowicz
Co-founder
Finance student turned co-founder. Runs finance, fundraising, strategy and operations: finding the capital, building the partnerships and making the decisions that keep the company moving.
- Full bio
- LinkedIn

Hamdaan Mohideen
Data Engineer
I’m a data engineer with three years of building pipelines that have to be right: batch and real-time, on Airflow, Kafka and Spark, including moving twenty million customer records onto a new platform. At Heuresia that means safety data where every record has a source and every result can be re-run. Before this I published on harmonising cardiac MRI data across scanner vendors, and built an exoplanet detector that reached the Microsoft Imagine Cup regional finals and won an Asia Pacific ICT Alliance award.
Sherwin Casem
Senior Data Engineer
Sherwin Casem is a Senior Data Engineer with over 12 years of experience designing and building scalable data platforms, ETL/ELT pipelines, and cloud-based data architectures. He specializes in Python, SQL, data integration, and data quality, with a strong focus on data versioning, lineage, and reproducibility. He is passionate about building reliable data infrastructure that empowers machine learning and scientific research.
Long Do Nguyen Hung
AI & ML Engineer
I’m an AI & ML Engineer at Heuresia, where we predict whether a drug might harm the liver before anyone even makes it. I work on evaluation and product: figuring out how to prove our models are trustworthy, and making sure their predictions are useful to the scientists who rely on them.
I came from fintech, and I brought a few habits with me: test against benchmarks you can’t game, make every result reproducible, and always ask how things could fail. In drug safety, just like in finance, trust has to be earned.
Alexandru Apetrei
Intern
Hello, I’m Alex! I’m passionate about technology, artificial intelligence, and building things that can make a real impact. I enjoy exploring new ideas, solving challenging problems, and continuously learning through hands-on projects and experimentation.
Ramcharan Velpuri
Intern
I’m an intern at Heuresia, learning how a drug-safety model gets built and proven from inside the team building it. I take on the work that needs doing around the engine, from checking data to keeping experiments tidy, and I ask a lot of questions along the way.
Amelia Sisik
Socials Manager
I run Heuresia’s social channels: how we talk about the engine, to whom, and when. That means turning benchmark results and toxicology into plain language that investors, partners and press can act on, without overclaiming. The science has to be honest, so the story has to be too.

Scientific advisors in toxicology, pharmacology and medicinal chemistry: a few engaged people rather than a wall of logos, named here as they join.

Why safety

Toxicity kills more drug candidates than anything else, and the verdict usually arrives late: after the chemistry, after the animal studies, after most of the money. A programme can survive a weak target or a slow route to synthesis. It rarely survives a liver signal in year four.

We chose that one call because it is the highest-value prediction in the pipeline, and because public safety data makes it the one we can prove on compute alone. No lab is needed to show the engine works. A frozen benchmark is.

01The highest-value callSafety decides whether a programme lives or dies, earlier than anything else about it.
02Provable on computePublic safety data means the first result is a model proven on a public benchmark, on data it never saw.
03Structure is enough to startThe engine reads a molecule’s structure and nothing else, so it can be asked before synthesis, while changing course is cheap.
04An honest fitA software team with university science access, aimed at one hard prediction, not a boil-the-ocean platform.

The long game

Prove the engine, then fund the drugs. Four stages, two of them underway.

Investor brief

The deck and the benchmark result, within two working days.

01Now
Data
Underway
Versioned, provenance-tracked safety data and a frozen benchmark, locked before any modeling.
02Now
Predict safety
Underway
The engine. Toxicity and drug-likeness from structure alone, with calibrated uncertainty on every call. Judged on a public benchmark, on data it never saw.
03Later
Design & make
Once the engine is proven
Generate our own candidates and keep only the ones that can actually be synthesized.
04Later
Validate
Once the engine is proven
Synthesis and assays through partners and CROs; measured results feed back to the engine.

How we work

01The benchmark is frozen before we train.Splits are leakage-free, by scaffold and by time, and locked before any modeling. A win has to survive outside scrutiny, or it is not a win.
02Every call carries its uncertainty.Each prediction comes with a calibrated confidence, so we know which ones to believe, and say so.
03Every record has a source.Datasets and results are versioned, traceable to where they came from, and re-runnable in one command.
04A toxicologist stands behind the number.Endpoints and labels are chosen and judged by people who know the biology, so a benchmark win means something in a clinic.

Part of

NVIDIA InceptionProgrammeNVIDIA’s programme for startups building with accelerated computing: technical resources, training and go-to-market support as the engine’s training runs scale.
altalabBuild programme, by AltaIR CapitalOne of the companies in the programme’s first cohort, Fall 2026. A programme, not an investment: we are part of it, not backed by it.
Medical University of WarsawFacultyToxicology and pharmacology faculty for defining endpoints and judging the data, and wet-lab access once the engine is ready for it.