FORMETISMachine learning you can read.
Formetis builds SymML: machine learning that returns a short algebraic formula instead of a black box. The formula is the model, so every prediction can be inspected, audited and defended.
Funded by
A credit risk model
risk = debt / income + 3 * missed_payments
It reads like a sentence: risk grows with debt relative to income, and every missed payment adds three points. No weights, no hidden layers. What you see is the whole model.
Three more examples below.
What SymML is
Most of today's machine learning is a black box. The models that drive decisions on loans, diagnoses and production lines contain millions of internal parameters. They can be very accurate, but nobody, including their developers, can say why a particular prediction came out the way it did.
Symbolic inference is the other way of doing machine learning. Instead of tuning a black box, it searches for a short algebraic formula that explains the data, and that formula is the model: there is nothing behind it to trust or to doubt. The idea is old, but it was held back by a hard problem. The space of possible formulas is astronomically large, and searching it was not computationally feasible beyond simple cases.
That is the problem our algorithm solves. Developed at KIT and implemented in our software SymML, it searches spaces of billions of candidate formulas on standard cloud GPUs, works on any tabular data, and is domain agnostic. It takes symbolic inference to a new level and makes applications possible that were not feasible before.
We don't just predict. We learn.
A black box gives you a score. A formula tells you what drives the outcome, and that is knowledge you can question, defend and act on. In regulated industries the formula is the audit. In science and engineering it is the discovery.
Materials discovery
stability = cohesive_energy / atomic_spacing**2
Read it: stability rises with how strongly the atoms bind, and falls sharply as they sit further apart. The model does not just rank candidate materials. It names the physical quantities that make a material stable, and a scientist can check that claim against known physics. That is what turns a screening run into a design rule, and it is how new materials get designed rather than found by trial and error.
Predictive maintenance
failure_risk = vibration * load * temperature**2
Read it: failure risk rises with vibration and load, and with the square of temperature, so heat matters twice over. That makes the model more than a warning system. Because it quantifies how strongly temperature drives failure, it also tells you what a better cooling system is worth. The same equation that predicts the failure justifies the investment that prevents it.
Genomics
disease_risk = gene_a * gene_b / gene_c
Read it: disease risk is driven by genes a and b acting together, held in check by gene c. In pharma an equation like this does three jobs. It points to targets, because blocking the interaction of a and b, or strengthening c, are concrete therapeutic directions. It defines the lab experiment, because knocking down one gene shows directly whether the formula holds. And it doubles as a biomarker, because measuring three genes is enough to find the patients most at risk, which is how a trial gets its inclusion criteria.
Like the credit model at the top of the page, these formulas are illustrations written for clarity. The ones SymML returns are learned from your data, and they read the same way.
Who it is for
SymML sells where decisions must be justified. Banks and insurers on credit scoring, medical diagnostics, industrial quality control: in these industries the buyers are the heads of model risk, data and compliance, who need to see why a model decides, test it against policy and sign off on it. From December 2027 the EU AI Act places binding transparency duties on high-risk machine-learning systems, and many companies will have to change how they build and validate their models. That change points toward interpretable AI, which is what Formetis sells: a market opening on a fixed schedule.
SymML also sells where the formula itself is the value. R&D teams in chemistry, materials, genomics and pharma use it to understand their data, not only to predict from it. Engineering teams use it in predictive maintenance, where knowing what drives a failure is worth as much as the warning.
Revenue comes from software licences and support, sold to businesses.
Open role
We are completing the founding team with a business co-founder in Berlin: go-to-market, fundraising, and the commercial build, with a double-digit equity stake. If that could be you, or you know the person it fits, contact us.
Founding team
We have worked together for years on interpretable machine learning for materials research, and we have seen what a readable model is worth. After many successful applications in that field, we are now bringing the approach to new industrial frontiers.
Dr. Luigi Sbailò · Co-founder
Luigi invented the algorithm behind SymML and leads its development. He is a machine-learning researcher at KIT with years of experience in scientific software development, high-performance computing and GPU deployment.
Dr. Lucas Foppa · Co-founder
Lucas connects SymML to industry and leads the pilot applications with our partners. He is a materials scientist at the Max Planck Society with years of industrial collaborations.
Prof. Luca Ghiringhelli · Scientific advisor
Luca is a professor at KIT and was among the earliest researchers to bring interpretable machine learning into the physical sciences. He advises Formetis on the science.
About the name
Formetis joins two words. Forma is Latin for form or shape, and it is the root of both formula and formalism. This is what we produce: equations, not just predictions. Metis is ancient Greek for practical intelligence, the resourceful kind that Homer gives Odysseus. Metis is also a goddess, the mother of Athena, and from her wisdom itself is born. Together the two words describe what we are after: practical answers in a form you can read.