Ammonix bets on specialized AI that acknowledges its limitations after Cortex’s billion-dollar sale

Following the sale of Cortex to Boston Scientific for up to $289 million, scientist Peter Ruppersberg launched Ammonix, a Swiss AI startup focused on compact models trained with specific data. Founded in spring 2025 in Blonay, a Swiss town on Lake Geneva, the company advocates for systems that can make quick decisions, explain their reasoning, and recognize when they lack sufficient information to respond.

Ruppersberg’s decade of experience treating cardiac arrhythmias taught him that in critical environments, a confidently delivered wrong answer can be more harmful than no answer at all. He notes that in medicine, being right most of the time is not enough.

Cortex, sold in January 2025 for $239 million in cash plus up to $50 million tied to clinical targets, developed OptiMap, a tool that maps the heart’s electrical activity to guide atrial fibrillation treatment. The acquisition was announced by Boston Scientific and detailed in documents filed with the SEC, the U.S. capital market regulator.

Ruppersberg, a former member of a Nobel Prize-winning research team, tenured professor, and author of articles in Nature and Science, also founded Mentasys, a pioneer in AI-powered product search, which was sold to Lycos in 2006.

From medicine to AI architecture

The sale of Cortex led Ruppersberg to question whether lessons from medical decision-making could inform a general architecture for AI systems. While the industry moves toward larger, centralized models, Ammonix is taking the opposite approach: creating small, specialized solutions that operate locally and can justify their decisions.

The startup’s architecture mirrors the sequence observed in cardiologists: rapid pattern recognition, decision-making, and explanation. Ammonix divides these tasks between two components: a fast, specialized decision-making module and a language model for explanation, questioning, and dialogue.

For Ruppersberg, the key lesson from clinical practice is that uncertainty is also a valid response. A responsible system must acknowledge a lack of evidence and refer cases to further analysis or human specialists. A fluent answer is not necessarily a good decision.

The company also prioritizes data quality over quantity. Hundreds of well-documented clinical cases can be more useful for specific decisions than millions of generic examples. The goal is to make an organization’s specialized knowledge more explicit and easier to share, rather than hiding it behind opaque models.

Business model and next steps

Ammonix operates with a small team and private funding, without external investors, and is focused on building its architecture before expanding commercially. Instead of selling access to a large language model like major labs do, the startup plans to generate revenue through commercial licenses, custom implementation, and system adjustments for each client.

The current challenge is teaching the machine to recognize the limits of its own knowledge and know when to refrain from answering.