In addition to widely used predictive models based on traditional clinical risk factors [1–2], some external facial features have also been linked to increased risk of CAD, opening up new opportunities for screening [3]. For example, features such as baldness, gray hair, facial wrinkles, earlobe crease, xanthelasma, and corneal arcus have been associated with higher risk of CAD and overall cardiovascular deterioration [4–6]. Several studies have also shown that such visual features may improve the accuracy of conventional risk models. Recently, similar research has been conducted in China and has shown impressive results [7].
Our goal: Development of an algorithm for detecting coronary artery disease risk based on facial photo analysis.
Features: clinical testing, open publications, scalable architecture.

Effective postoperative monitoring is a crucial stage in modern patient care. Timely detection of complications or negative trends allows for rapid treatment adjustments, prevents severe consequences, and reduces rehospitalization risk. However, traditional follow-up by phone is labor-intensive for medical staff, especially with large patient numbers. Implementing an AI assistant capable of independently calling patients, collecting key health parameters, and promptly advising on seeking medical care when needed greatly reduces the physician's workload [1-2]. Such an AI assistant works strictly according to protocol, never forgets to ask critical questions, and can even detect hidden risks based on patient history. This enables clinicians to focus on more complex cases, improve the quality and personalization of care, and save significant time, reducing the load on medical personnel [3].

Studying the properties of matter and generating matter with given properties is the basis for many areas of industry. Neural networks have recently begun to be actively used in the field of materials science and this is one of the hottest scientific areas [1]. Large language models (LLMs) have shown impressive abilities in a variety of tasks, but there are still challenges in their use for natural sciences [2]. In our project, we aim to adapt and explore the behavior of LLMs for generating atom-level structures for solving physical and chemical problems. A major challenge in generative modeling is the limited availability of (especially experimental) data [3]. In addition, the choice of data representation for materials is not obvious and depends on the approach used. With some representations, a portion of the object's information is lost. We adapt and extend textual representation methods to low-dimensional crystal structures, and for molecular structures, we propose a multimodal representation to improve generation quality. The ultimate goal is to develop algorithms for determining the composition and geometry of crystals and molecules, accounting for their physically important macroscopic properties. Research is ongoing [4-6], and many generation methods have been proposed, but most have important limitations. For crystals, particular attention should be paid to optical properties and low-dimensional (2D) materials, which are still understudied and whose number grows every year. 2D materials [7] are highly promising for applications due to their unique properties and size, and they are relatively easy to construct experimentally.

Coronary artery disease (CAD) remains a leading cause of death and morbidity worldwide, including in Russia, where it accounts for over half of all cardiovascular diseases. Despite progress in medicine, effective CAD diagnosis and treatment are still a top healthcare priority. Modern imaging techniques such as coronary angiography provide a wealth of information, but the growing volume of data requires significant effort to process and interpret, and human factors can affect diagnostic accuracy and lead to extra risks and costs. Developing automated tools to rapidly and accurately analyze coronary angiography results and calculate key blood flow parameters in real time can improve diagnostic quality and reduce strain on clinicians and healthcare systems. This is now possible due to advances in machine learning and artificial intelligence, which have already demonstrated success in medical image analysis [1]. The project involves building a multi-stage platform for comprehensive coronary artery assessment, including artery segmentation [2], 3D modeling [3-5], hemodynamics modeling, and automated ischemic event risk assessment based on these data. This comprehensive approach will not only increase CAD diagnostic accuracy, but also optimize treatment costs, making it highly relevant and practical for the medicine of the future.