The modern dynamic production environment requires
continuous improvement of tools for managerial
decision-making. Although decision-making is increasingly
directed toward artificial intelligence, the most important
managerial decisions are still related to solving
combinatorial optimization problems. Numerous papers
presenting improvements to existing procedures are
published daily; however, the implementation of the
proposed algorithms in practice is often unclear and
complicated. In addition, experimental evaluations of these
algorithms, even in the most reputable journals, often
contain bias, which makes it difficult for potential users
to select and apply an appropriate algorithm.
This lecture aims to present a new, publicly accessible
environment—the web application OPR-MAN (Angular, .NET 8,
SQL Server)—which enables users to compare, apply, and
visually present the results of the best algorithms for the
most important problems in operations management:
sequencing and allocation problems, clustering problems,
routing and packing problems, as well as material
requirements planning problems. For each module, users can
upload benchmark problems from the literature or create
their own instances, solve them, and compare the obtained
solutions with those generated by other algorithms for the
same problem. The visualization of solutions is graphically
rich; for example, routes are displayed on real maps,
packing solutions are graphically illustrated, and similar
visual representations are provided for other problems.
The application has already found practical use in many
companies within supply chains. A particularly important
application of this system is in education, as it allows
students to follow every step of the algorithmic solution
process. The application is used in the training of
students at the Faculty of Organizational Sciences, where
it has been evaluated as a significant contribution to the
development of competencies of future operations managers.

While deploying models on device is still the ideal
approach for mobile edge AI, resource limitations often
prevent large neural networks from running locally. In
these cases, streaming multimodal sensor inputs to a remote
inference service provides a practical path to deliver
interactive AI functionality. In this context, low-latency,
bidirectional media transport becomes a first-order
requirement for real-time human with machine interaction.
This paper examines WebRTC (Web Real-Time Communication) as
a communication substrate for such AI offloading, enabling
sub-second streaming of multimodal inputs (audio/video) and
synchronous responses without the overhead of polling-based
architectures. We describe the protocol stack end-to-end,
including application-level signaling (SDP offer/answer and
trickle ICE), NAT traversal via ICE with STUN/TURN, and
mandatory security through DTLS-SRTP. We further discuss
multi-party topologies and analyze Selective Forwarding
Unit (SFU) architectures as a practical scaling strategy
for interactive, AI-assisted conferencing. Overall, WebRTC
emerges as a key enabler for the next generation of
intelligent mobile systems where real-time perception and
response are achieved through remote inference rather than
on device.

Interlanguage full-fledged compatibility with C++ is
still one of the open tasks of modern programming. Its
complexity
lies in assembling the necessary data to use the
functionality.
Traditional methods use wrappers to assemble this data, but
this
method is limited in use. Since most of the necessary data
in
different versions is in a binary file, and recent research
has shown
that AI is effectively reduced with binary code analysis, a
new AI
model has been developed to solve the problem. However, it
is
known that AI may not give accurate answers and/or generate
them. This article discusses the process of learning a
model for
binary analysis with an emphasis on rigorous search. The
tests
show that, as a result of training, the model provides a
high fineness
of answers compared to known solutions, even with incorrect
answers.

This paper will present the details of the construction of
the author's new chess program Axon II (version 2026). The
combinatorial explosion problem, which is fundamental to
all applications for intensive decision tree processing, is
addressed in the new engine by a combination of modern tree
processing techniques and intensive use of X86 machine code
for programming the basic functions of the engine,
primarily the position generator and evaluator. Empirical
tests on various platforms show very high raw engine speed
factor (NPS factor) in comparison with referent chess
engine - Stockfish 18.

This paper proposes a two-stage framework for hate speech
detection and categorization in emoji-containing text. The
framework first performs binary hate speech detection and
then assigns a fine-grained hate category only to texts
predicted as hateful. To examine the effect of emojis, the
study uses paired texts in two forms: an emoji-containing
version and a corresponding emoji-free version. The
experiments were conducted on a dataset with paired samples
and evaluated across several locally hosted large language
models. The results show that binary detection is generally
more stable and more accurate than fine-grained
categorization. They also indicate that emoji removal has a
relatively limited effect on hate speech detection, but a
stronger effect on category assignment, suggesting that
emojis influence the semantic interpretation of hateful
content more than its initial recognition. Among the
evaluated models, qwen3 achieved the best overall
performance, while phi4 provided the most balanced
trade-off between effectiveness and efficiency. The
findings highlight the importance of emoji-aware analysis
and show that a two-stage approach offers a useful
framework for studying both the presence and the type of
hate speech in social media text.

Evaluating End-User Development (EUD) authoring tools
requires empirical studies in which participants complete
representative tasks. In many domains, such tasks
necessarily involve producing a complete artifact composed
of multiple distinct elements. Measuring only whether the
final artifact is correct offers limited insight into where
and why users fail, and does not support meaningful
comparison across tools that differ in how they approach
the authoring process. To address this limitation, we
propose a protocol for deriving domain-specific units of
analysis, which are artifact elements that correspond to
relevant domain concerns and can be assessed for
correctness independently of any specific tool's notation
or authoring technique. The protocol consists of four
phases: Artifact Corpus Construction, Individual
Identification, Consolidation and Refinement, and Coverage
Validation, with domain experts involved in the latter
three. We demonstrate the execution of the protocol through
its application to the domain of smart contract authoring
tools, where a representative task cannot be smaller than
creating a complete contract and existing tools differ
considerably in how they support that process. The absence
of a systematically derived set of such units in prior
literature provided additional motivation for this choice
of domain. This paper focuses on the protocol and reports
its implementation in the smart contract authoring domain,
while a detailed analysis of the resulting units is
reserved for a subsequent study.