Artificial intelligence is a working tool based on a balanced combination of algorithms and engineering. Experts and doctoral students from the HSE Moscow Institute of Electronics and Mathematics explain how AI technologies can improve an application, device, or system, and what engineering tasks are solved in the process.
Denis Korolyov, Head of the HSE MIEM Laboratory of Network Videotechnologies
AI as a Teacher’s Assistant
We have created an LLM-based ‘evaluator’ for students’ open answers: the student receives not just a grade but also a detailed explanation behind it. The goal is to provide quick feedback during lectures, ensure objective assessment in exams, and collect a digital learning footprint. The solution is that the LLM evaluates answers based on completeness and correctness criteria. If the student disagrees with the assessment, the dialogue is sent to the teacher for manual verification.
Selecting the right tools and ensuring validation pose an engineering challenge: the system asks questions based on thematic proximity. When a correct answer is received, it offers either closely related or more distant (within the section) follow-up questions. The experiment showed 90% agreement with expert assessment within one step of the grading scale.
The same tool can be used to collect feedback. Another of our teams is analysing SET (student evaluation of teaching) comments, with one of the challenges being the abundance of abbreviated responses like ‘everything is OK’ or ‘it’s bad.’ The LLM-based survey tool helps to conduct a conversation by asking follow-up questions. We have recently launched the feature of asking additional questions in the exam mode.
Andrey Vasenko, Deputy Head of the HSE MIEM Quantum Nanoelectronics Laboratory
The ‘Smart Accelerator’ of Physical Modelling
We study perovskite solar cells, where structural defects reduce efficiency by trapping charge carriers and converting their energy into heat. Precise quantum mechanical calculations can identify the role of each defect and suggest ways to counter them. However, true quantum calculations, even for several hundred atoms, require immense computational power.
We solve this problem by using neural network models trained on quantum computing data: they predict forces and energies with almost quantum-level accuracy but thousands of times faster. This makes it possible to model systems of tens of thousands of atoms and observe the dynamics of defects (ion migration, clustering, and the evolution of grain boundaries).
This approach sets the following engineering tasks: to identify the most informative geometric descriptors of defects by analysing mutual information, to develop an ML model for predicting the quantum dynamics of charge carriers, and to build an end-to-end pipeline from atomic physics to applied device parameters (photoluminescent response, voltage losses, and stability during thermal cycling and radiation exposure).
Sergey Slastnikov, Head of the HSE MIEM Unit for Applied AI
Fine-Tuning of Models with Limited Resources
Working with AI solves a wide range of engineering problems. We were approached by colleagues from the ANO Nauka i Obrazovaniye (Penza) with a proposal to develop a preventive service that would show teenagers what their appearance would look like in 20 years if they maintained bad habits. The task proved challenging—primarily because we lacked suitable visual data to train the model to transform individual features: collecting such a dataset is difficult, especially with a focus on adolescents.
Our research focused on open data and models, and the organisation provided cloud computing resources. As a result, we applied low-rank adaptation technology (LoRA): we used a small dataset to fine-tune a large pre-trained model, redirecting its general knowledge to solving a narrow task. In this way, we managed to maintain quality with limited resources.
Assessing the plausibility of the predictions was a separate engineering challenge. Testing became part of the engineering workflow: we created a representative sample and evaluation criteria, and deployed the prototype on a dedicated server. Students then tackled applied engineering tasks ranging from data collection and cleansing to result validation and deployment.
Ilya Semichasnov, Head of the HSE MIEM Game Engineering and Interactive Systems Workshop
AI as the Basis of a Vibe-Coding Engine
Our Claw Engine enables fast prototyping of individual elements of games and interactive applications. At first, we tried to generate all content using our own AI models—from sound assets to 3D models and textures. We achieved some results in texture generation using diffusion models, but we quickly abandoned total generation due to a lack of high-quality datasets, prohibitive training costs, and the fact that many game assets were already available in public libraries. Our current strategy is to competently integrate readily available solutions to achieve automation at different stages of development. Claw Engine aims to become a vibe-coding engine: the user describes the game concept as plain text, and the system assembles it from readily available components.
This requires multi-layered engineering: action orchestration (an interface between the neural network and the engine, similar to the MCP connector); artifact control (filters to screen out solutions that would be nonsensical in gameplay); context retention during iterative refinement; and fine-tuning the assembly pipeline to ensure that the generated code works correctly with physics, animations, and assets. Separately, we are developing the management of non-player characters based on reinforcement learning—an approach promising for games and robotics but extremely demanding in terms of computing power.
Felix Smirnov, doctoral student at the Doctoral School of Engineering Science, research assistant at the HSE MIEM International Laboratory for Supercomputer Atomistic Modelling and Multi-Scale Analysis
AI as an Object of Research in Supercomputer Communications
A key task is to increase the speed of distributed LLM training by increasing the performance of collective GPU communications across several network stages. The goal is to optimise LLM workloads and save time and computing resources on the supercomputer, thereby making AI more efficient in industrial and business applications.
The solution is to establish a comprehensive method for analysing training performance: from the macro level (general application behaviour and node-to-node data exchange) to the micro level (performance and resource usage at the level of individual GPU cores).
The engineering novelty lies in the model of a hardware-software communication stack, where each layer contributes to the final performance of collective operations. The tools used are HPCToolkit and Score‑P to collect metrics and profiles; HPCViewer, Vampir, and Cube for visualising time charts and detecting bottlenecks; and network microbenchmarks (OSU, ReproMPI) for estimating communication delays. The object and environment of the study is the supercomputer itself—dozens of nodes with multiple GPUs, where we test node-to-node interaction when transferring large amounts of data. The results have been presented at specialised conferences (the International Conference on Parallel Processing & Applied Mathematics, Russian Supercomputing Days, and Parallel Computational Technologies) and published by Springer and in IJHPCA.