12th edition

CAMAI 2026

20 – 21 November 2026 Hybrid, Bucharest, Romania
The history of computer science seminar is dedicated to Grace Murray Hopper, who built the first compiler and made programming languages readable.
Portrait of Grace Murray Hopper in United States Navy uniform
Grace Murray Hopper, 1906–1992. Read her biography
0
Attendees
0
Days
0
Talks
0
Practical Applications

About CAMAI

The Conference on Applied Mathematics and Informatics is organized by Faculty of Engineering and Informatics from Spiru Haret University and by Research Center in Mathematics and Informatics, affiliated to Central Institute for Scientific Research from Spiru Haret University.

The conference is mainly addressed to researchers in mathematics, computer science and information technologies, and includes theoretical and applied mathematics and computing topics.

The conference takes place in Bucharest, 46Gth Fabricii Street, the location of Faculty. Every year, the conference is dedicated to a major local, national or international event.

The mission of Cybersecurity 2030: Are We Prepared for the Next Decade of Cyber Threats? extends beyond cybersecurity, embracing the interconnected disciplines of software engineering, computational methods, artificial intelligence, and machine learning. This event seeks to provide a multidisciplinary platform where experts, researchers, and practitioners come together to address the complex challenges that span across these fields. By integrating insights from software engineering, computational modeling, and the transformative power of AI and machine learning, the conference aims to drive innovative solutions that enhance the resilience and adaptability of cyber defense mechanisms.

Participants will explore how these fields converge to fortify cybersecurity, from developing robust software architectures to deploying machine learning models capable of real-time threat detection. The event emphasizes the essential role of interdisciplinary collaboration in preparing for the next decade of digital challenges, highlighting how advances in software and computational methods power critical cybersecurity initiatives and how AI accelerates threat prediction, analysis, and response.

Through this mission, CAMAI 2026 aspires to build a collaborative community of forward-thinking professionals dedicated to safeguarding the digital landscape, fostering innovation at the intersection of cybersecurity and emerging technologies, and nurturing the next generation of leaders in engineering, data science, and cyber defense.

In an era where digital threats continue to evolve in complexity, quantum computing and quantum cryptography are set to revolutionize the cybersecurity landscape. The mission of CAMAI 2026 recognizes the transformative potential of these fields, especially as traditional cryptographic methods face increasing vulnerability in the wake of quantum advancements. Quantum computing promises unprecedented computational power that could both advance cybersecurity defenses and, conversely, pose significant risks by rendering conventional encryption methods obsolete.

The conference seeks to highlight the critical importance of quantum-resistant cryptographic solutions and the development of protocols that can withstand the capabilities of quantum-based attacks. By exploring quantum computing’s impact across software engineering, AI, and cybersecurity, CAMAI 2026 aims to prepare participants for the next frontier in secure computing. Embracing quantum cryptography’s promise of unbreakable encryption, the event will delve into current research and practical strategies for integrating quantum-safe measures, underscoring the urgent need for preparedness in an approaching quantum era. This focus not only elevates the importance of interdisciplinary collaboration but also places CAMAI 2026 at the forefront of innovation in secure digital technologies.

The organizers wish this conference to be an important international event in applied mathematics and computing.

The first edition occurs in 2007.

From 2013, the conference includes a special section for History and Pedagogy of Mathematics and Computer Science. In 2026 that section is dedicated to Grace Hopper.

This edition is dedicated to Grace Hopper

Grace Murray Hopper took a PhD in mathematics at Yale in 1934, joined the Navy Reserve in 1943, and was assigned to the Bureau of Ordnance Computation Project at Harvard, where she programmed the Mark I. In 1952 she wrote the A-0 system, the first compiler, at a time when the prevailing view was that a computer could only do arithmetic. Her work on FLOW-MATIC led directly to COBOL, and to the idea that a program could be written in something close to English and then translated for the machine.

She also spent decades teaching. She handed out lengths of wire cut to 11.8 inches, the distance light travels in a nanosecond, so that an audience could hold the cost of a delay in their hands. The history and pedagogy of computer science session at CAMAI 2026 takes that approach as its subject: how the field explains itself, and how it teaches the next generation.

  • 1934PhD in mathematics, Yale
  • 1944Programmed the Harvard Mark I
  • 1952Wrote the A-0 compiler
  • COBOLStill running in production

Why Should I Participate in CAMAI?

  • Unique chance to find passionate professors and professionals about their fields
  • Learn from the best in the academia and industry
  • Discover the best tools and practices in different fields of computer science and technologies
  • Meet scientisits, researchers and developers from all over the world
  • Grow your network and gain knowledge
  • Practical workshops with excitment simulations and practical demonstrations
  • After-party

Schedule

Academic & Industry Section


10:00 - 10:30

CAMAI 2026 Organizing Committee

CAMAI 2026 Welcome

On-line




10:30 - 10:50

Cornel Argint

The electromagnetic spectrum in combating drones

On-line
The armed conflict in Ukraine has brought a new paradigm in the planning and conduct of military operations - drones have become an increasingly important vector for the achievement of military objectives. An essential factor in their use in high-precision actions is represented by the use of electromagnetic waves. In this way, another field of confrontation appears in the electromagnetic environment, leading to new challenges in the field of electronic warfare.

Authors: Cornel Argint1,2
Affiliation:
1National Cyber Security Directorate (DNSC)
2Faculty of Engineering and Computer Science, SPIRU HARET University

10:50 - 11:10

Rodica Ioan
History of computer science

Grace Hopper – the pioneer of modern programming

On-line
This presentation explores the life and contributions of Grace Hopper, a pioneer of modern computer science. It covers her achievements in the development of compilers and the COBOL language, as well as her influence on the evolution of programming. Through intelligence, courage, and perseverance, Grace Hopper blazed new trails in the world of technology.

Authors: Grigore ALBEANU, Rodica IOAN
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

11:10 - 11:30

Valerica Moșneguțu

Elie Carafoli, founder of journals

On-line
Elie Carafoli founded the Chair of Aerodynamics and Fluid Mechanics within the Electromechanics Department at the Polytechnic School of Bucharest in 1928—was also the founder and first Editor-in-Chief of the journal ‘‘Revue Roumaine des Sciences Techniques. Série de Mécanique Appliquée” in 1956. This world-renowned professor emeritus also founded the journal Studies and Research in Applied Mechanics, in 1950.

Authors: Valerica Moșneguțu
Affiliation: Institute of Solid Mechanics, Romanian Academy, Bucharest

11:30 - 11:50

Rodica Ioan

A new computational approach with application to the population modeling

On-line
The application of the scaling laws to classification of the phenomena in the context of the time-dependent predator-prey systems is presented in this paper. The prey has a life history of immature and mature stages. The interaction between the prey and the predator is based on the energy flow from one organism to the other. The predator is defined as an organism that primarily obtains food by killing and eating other organisms. Predators control the prey population by maintaining the balance of the food

Authors: Veturia CHIROIU1, Ligia MUNTEANU1, Rodica IOAN1,2, Valeria MOȘNEGUTU1
Affiliation:
1Institute of Solid Mechanics, Romanian Academy, Bucharest
2Faculty of Engineering and Computer Science, SPIRU HARET University

11:50 - 12:10

Iulian Girip

Linear analisys of a thin circular plate with central load and simply supported edges

On-line
A thin circular plate with central load and simply supported edges is analyzed here. More precisely, a plate linear analysis is performed using Kirchhoff-Love theory for thin plates and assuming small displacements and linear elasticity. The analytical results are verified using FEA SolidWorks simulation. This thin circular plate study is part of a more extensive further study concerning a pack of thin circular plates (thin disks pack), with the circular plates having a hole in the middle. Subjected to bending by using a central load and with the edges simply supported on the other side, such a thin disks pack represents a very simple and efficient elastic damping element. The effective design of such elastic damping elements is based on the correct understanding of the bending and friction phenomena between the thin disks of the pack.

Authors: Iulian GIRIP1, Dan N. DUMITRIU1, Ruxandra Diana ILIE2, Marius IONESCU1, Valerica MOŞNEGUŢU1, Cristian RUGINĂ1
Affiliation:
1Institute of Solid Mechanics, Romanian Academy, Bucharest
2Technical University of Civil Engineering of Bucharest

12:10 - 12:30

Cristian Liviu Stefan

From Loving Grace to the DarkFace

On-line
The current state-of-the-art of agentic AI in the realm of e-safety involves challenges that emphasize the recently-acquired abilities, like prompt injection, out-of-bounds attacks (sandbox evasion) and alignment misconfigurations. From InkJect to self-replicating AI, this presentation aims to assess today’s cybersecurity landscape in order to better understand what the AI-driven future will look like and how to timely adapt.

Authors: Cristian Liviu Stefan, Nicoleta Iacob
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

12:30 - 12:50

Valentina Marascu

A proof-of-concept method for integrated cold-plasma diagnostics and secure data management

On-line
Cold-plasma experiments produce heterogeneous measurements that complicate diagnostic analysis and secure data exchange. This paper proposes a proof-of-concept method integrating Langmuir probe signals and optical emission spectra within a unified acquisition, storage, and processing framework. A structured data model supports synchronization, provenance, and consistent interpretation across instruments. Physics-informed analysis incorporates quasi-neutrality, current-voltage characteristics, and Boltzmann-plot consistency to constrain plasma parameter estimation. The framework is designed for limited labeled experimental datasets. For secure collaboration, time-resolved plasma observables provide deterministic input to cryptographic key derivation and update authentication. The method is examined using hollow-cathode plasma datasets for helium, hydrogen, and deuterium. Initial security tests confirm reproducible key generation, successful encryption and decryption, and normalized inter-source Hamming distances. These results demonstrate functional integration of diagnostic data handling and plasma-anchored security. Experimental assessment of diagnostic accuracy, source min-entropy, and multi-laboratory deployment remains necessary before broader applicability or security guarantees can be established.

Authors: Valentina Marascu, Marius Iulian Mihailescu and Stefania Loredana Nita
Affiliation: National Institute for Laser, Plasma & Radiation Physics (INFLPR), Faculty of Engineering and Computer Science, SPIRU HARET University, Institute of Computers

12:50 - 13:10

Iulia Ciordas

Forecasting Stock Index Volatility with a Hybrid GARCH–LSTM Residual Model: A Stochastic Volatility Perspective

On-line
Volatility is a central quantity in mathematical finance: it drives option prices, risk measures and portfolio
decisions. Unlike returns, volatility is persistent and partly predictable, which makes it a more realistic forecasting target than the price itself. Classical models such as GARCH describe volatility clustering with only a few parameters, but their fixed functional form may miss nonlinear effects. LSTM networks are
more flexible, but they are hard to interpret and can overfit noisy financial data. In this paper we study a hybrid model in which a GARCH(1,1) forecast is used as an analytical baseline and an LSTM network is trained to predict the residual between this baseline and a range-based volatility proxy. We motivate the
setting through stochastic volatility models and the Feynman-Kac representation of option prices, which shows why the variance process is the key state variable. We compare the hybrid model with a naive forecast, GARCH(1,1), HAR and a standalone LSTM on daily data of the [S&P 500 index, 2005–2025], using
a walk-forward evaluation, RMSE, MAE, QLIKE and Diebold Mariano tests. We also analyze calm and turbulent market periods separately. This work is a first step toward combining classical financial mathematics with machine learning for time series forecasting, and it can serve as a starting point for further research on hybrid volatility models.

Authors: Iulia Ciordas
Affiliation: Polytehnic University Timisoara

13:10 - 13:30

Andreas Poletti

Secrets Without Borders: Securing Multi-Cloud Workloads with HashiCorp Vault

On-line
This presentation explores identity-based secrets management in multi-cloud environments using HashiCorp Vault. Through practical integrations with Microsoft Entra ID, GitLab, Kubernetes, Azure, and AWS, it demonstrates how centralized authentication, fine-grained authorization, and short-lived credentials can reduce security risks and strengthen infrastructure resilience.

Authors: Andreas Poletti
Affiliation: Poletti Technologies

13:30 - 13:50

Silviu Calota

Redundancy Is Not Backup: Building a Resilient IT Infrastructure

On-line
A redundant infrastructure is not necessarily a protected one. Drawing on real-world data center scenarios, we will explore the role of redundancy, virtualization, and backup, identify common single points of failure, and discuss how to design an IT infrastructure where a technical incident does not turn into a disaster.

Authors: Silviu Calota
Affiliation: Baneasa Business Development

13:50 - 14:10

Florin Postolache

Physics-Informed Maritime Cyber Range for Cyber-Physical Security Assessment

On-line
Modern ships rely more and more on connected IT and OT systems. Because of this, a cyber incident doesn't stay on the network anymore. It can change how the vessel actually behaves. Most existing cyber ranges focus on detecting and responding to attacks at the network level and say little about what happens to the physical plant when an intrusion succeeds. In this paper we describe the design, implementation and experimental validation of a physics-informed Maritime Cyber Range. It combines network segmentation, industrial control systems and dynamic simulation models in a reproducible cyber-physical testbed. The architecture is split into isolated functional zones that match the main shipboard systems: navigation, propulsion, cargo, communications and administration. It follows the cybersecurity principles set out in IACS Unified Requirement E26. The platform brings together PLC-based automation, human-machine interfaces (HMIs), simulated propulsion and electrical power systems, and network security monitoring. Its main contribution is that each attack scenario is tied to a physical response we can measure. This lets us study how malicious control actions lead to operational anomalies and how alarms spread through the system. As a validation case, we simulated an unauthorized override of the engine command. Shaft speed dropped, and the change showed up as indications in both the engine control room and on the bridge. We also used fault injection and negative testing to see how the system behaves when things go wrong and to find implementation weaknesses that standard functional tests tend to miss. The results show that a reproducible maritime testbed can link network and control-system events to consequences that operators can actually see. The range can be used to evaluate maritime security controls, run realistic attack-defense exercises and study operational resilience under controlled conditions. Next, we plan to validate full IT-to-OT attack chains, add threat scenarios against navigation systems and support multi-user training. Overall, this work offers an experimentally tested architecture for maritime cyber resilience research, so cyber-physical vulnerabilities can be assessed systematically without putting real vessels at risk.

Authors: Florin Postolache, Teodor Petrut Ilie
Affiliation: "Mircea cel Batran" Naval Academy, Faculty of Marine Engineering

14:10 - 14:30

Lunch Break

Networking break.

Hybrid Section: Students, Industry and Academic


14:30 - 14:45

Nataliya Lozanova

Recursive Sorting Algorithms in Haskell: A Functional Programming Approach

On-line
This paper presents the implementation of five classic sorting algorithms (QuickSort, MergeSort, InsertionSort, SelectionSort and BubbleSort) in Haskell, a purely functional programming language. Since Haskell has no loops and no mutable variables, every algorithm is expressed through recursion, pattern matching and higher-order functions such as filter and foldr. All implementations are generic over the Ord type class, so they can sort any comparable data type. We compare the algorithms in terms of time complexity and of how naturally each one fits the functional style. Divide-and-conquer algorithms such as QuickSort and MergeSort translate almost directly from their mathematical definitions, while algorithms based on swapping elements in place, such as BubbleSort, become less natural because each step must build a new list. The implementations are tested by treating the sorting functions as values and checking each result with a generic sortedness predicate. Future work includes property-based testing with QuickCheck and performance benchmarks on large inputs.

Authors: Nataliya Lozanova
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

14:45 - 15:00

Andrei Constantinescu

An Interactive Temperature Converter Architecture in Haskell: A Functional Paradigm Approach

On-line
This paper introduces the structural architecture and design principles of an interactive, multi-scale temperature converter developed entirely within the pure functional paradigm using Haskell. Traditional software implementations of everyday utility tools often rely heavily on imperative structures, introducing mutable states and prone-to-error loop variations. Our approach mitigates these structural inefficiencies by enforcing strict mathematical function purity and leveraging Haskell’s native type-safety mechanism to handle thermodynamic transformations between Celsius, Fahrenheit, and Kelvin scales seamlessly.

Authors: Andrei Constantinescu
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

15:00 - 15:15

Ionut Bara

ForensIQra: Design of a Web-Based Framework for Blockchain Transaction Analysis and Digital Forensics

On-line
The increasing adoption of blockchain technology and cryptocurrencies has introduced new challenges in cybersecurity and digital forensic investigations. Although blockchain transactions are publicly recorded, analyzing transaction histories and identifying relationships between wallet addresses can be complex and time-consuming.
This paper presents the proposed design of ForensIQra, a web-based framework intended to support blockchain transaction analysis and digital forensic investigations. The system aims to facilitate the retrieval, processing, and visualization of publicly available Ethereum blockchain data.
The proposed architecture includes a Python-based backend developed using FastAPI, blockchain data integration, and a web interface designed to display transaction histories, address relationships, and relevant transaction information.
The project also explores the potential of graphical representations and structured reports to improve the interpretation of blockchain transaction flows.
ForensIQra is intended to serve as an educational and investigative support tool, providing a foundation for further development in cryptocurrency transaction analysis and digital forensics.

Authors: Ionut Bara
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

15:15 - 15:30

Vlad Poienariu

Midna - Server monitoring and alerting tool

On-line
Midna is a self-hosted solution for small servers and homelabs, this also being its key advantage over popular existing market solutions such as Grafana or Prometheus. While they are also self-hosted, another advantage is the very low complexity and resource usage of the Midna app.

The idea came from a real problem I was facing while administering my personal homelab, that problem being that I didn't know when something broke, when something went down, and so on. The only way for me to know this was to SSH into the server, which takes time and requires a VPN to access my LAN (Local Area Network). So I decided to approach this exact problem I was having by creating a very lightweight and easy-to-use and upgrade app that has only one scope: notify whenever something goes wrong, and this is exactly what I did.

Basically, we have an app that can monitor the activity of multiple websites, monitoring their status and sending an alert if the status received is different than 200 (A status other than 200 means something is not OK, maybe a client issue or server issue, but there is a problem). It monitors CPU usage, CPU temperature, RAM memory usage, and also ROM storage, sending an alert if the current free storage is below a certain threshold. These alerts work on a simple principle: if the current value is inside a threshold or above a value, trigger an alert. These thresholds and values can easily be modified by the server administrator by editing the config.cfg file, which has all of these items inside. Furthermore, these alerts are not sent constantly if, for example, CPU utilization stays at 100% for a few minutes. Only one alarm will be sent once every X minutes, X being a value the server administrator can also change from the config file, this leading to a more precise and dynamic way of monitoring things based on the usage baseline already known.

Alarms, also referred to as notifications, are sent using a third-party app called Ntfy. The server administrator must only create an account and create a topic on their official website, then paste the topic name inside the config file, and the alerts should be received properly. I chose Ntfy for this because of the main principle I developed the app itself around: simplicity. You do not need an API key or any other form of preconfiguration to use it; you just need to create a topic. Indeed, this is not secure because anyone can see those notifications, but this is not the scope of this app. The scope is for a private homelab or a small server agency/user to monitor their device easily. This can also be improved in the future by using a private topic accessible only via a password or API key.

Besides the general features and possibilities of the Midna app, it can also monitor backup logs and check for a specific line that indicates whether the backup has completed successfully or not. Using a cron job, we can run a specific script that will start at the same time as the backup and will check every few minutes for the backup completion. When it finds the line that indicates the backup was successful, it will send a notification about this, or, if the timer expires, it will send a notification that the timer failed or took too long.

Authors: Vlad Poienariu
Affiliation: Faculty of Psychology and Education Sciences, SPIRU HARET University

15:30 - 15:45

Stefania Loredana Nita

Computing on Encrypted Locations: A Survey of Homomorphic Encryption for Geospatial Data

On-line
Location data is among the most revealing categories of personal information. Trajectories and points of interest can expose home and work addresses, health conditions, and social relationships. As spatial datasets are increasingly outsourced to cloud platforms, data owners must choose between confidentiality and the benefits of third-party computation. Homomorphic encryption (HE) removes this trade-off in principle by allowing computation directly on ciphertexts. This paper surveys the use of HE for geospatial data processing and classifies existing work by the spatial operations it supports: distance computation, proximity and range queries, k-nearest-neighbour search, point-in-polygon tests, and aggregate analytics over spatial grids. It compares partially homomorphic approaches with the BFV, BGV, CKKS, and TFHE schemes, focusing on how each handles the arithmetic and comparisons that spatial algorithms require, and on the trade-off between exact integer encodings and approximate arithmetic for coordinates. The main recurring obstacles are the cost of non-linear operations, ciphertext expansion, noise management, and the lack of standardized benchmarks. Since the leading HE schemes rely on lattice problems believed to resist quantum attacks, the paper also discusses their role in the long-term protection of location data. It concludes with research directions, including hybrid designs with searchable encryption or trusted execution environments, applications in GNSS integrity and transport logistics, and reproducible evaluation frameworks.

Authors: Stefania Loredana Nita, Marius Iulian Mihailescu and Valentina Marascu
Affiliation: Institute of Computers, Faculty of Engineering and Computer Science, SPIRU HARET University, National Institute for Laser, Plasma & Radiation Physics (INFLPR)

15:45 - 16:00

Rares Muntean

Deep Autoencoder Framework for Real-Time Threat Detection and Dynamic Firewall Mitigation

On-line
Traditional network security perimeters face severe challenges when addressing zero-day exploits, major DDoS attacks, and stealthy internal network threats. While deep learning offers promising capabilities for anomaly detection, existing solutions face three practical bottlenecks.

1. The Domain Gap. Models pre-trained with static public datasets (such as CIC-IDS-2018) fail to generalize across diverse enterprise architectures, generating false-positive results;

2. Encrypted Tunnel Blindness. Standard physical network sniffing cannot inspect inside encrypted VPN tunnels, leaving private enterprise communications unmonitored;

3. Passive Remediation Delays. Conventional Intrusion Detection System (IDS) only issue passive alerts, allowing malicious traffic to reach backend services before human administrators can intervene.

In this paper, we present an autonomous, self-calibrating network defense pipeline that unites unsupervised behavioral analysis and automated kernel-level traffic filtering.

The packet capture module is designed to listen to both physical network cards (such as eth0) and virtual VPN adapters (such as tun0 or wg0). Capturing from virtual interfaces allows the system to analyze decrypted traffic moving through private VPN tunnels before it reaches local applications. The system also groups incoming and outgoing traffic into bidirectional flows. For each completed flow, the framework uses 13 key statistical features, including logarithmic packet inter-arrival times (IAT), byte ratios between client and server, port numbers and TCP control flags.

To avoid the need for massive datasets, the system allows the user to record a local “safe baseline” during normal, benign usage. To keep the model from memorizing exact packet sizes and delays, data augmentation is applied during training by adding artificial noise (random variations in timings and packet lengths). During live monitoring, the model reconstructs incoming flow features; if the Mean Squared Error (MSE) reconstruction loss exceeds a calibrated threshold, the flow is flagged as an anomaly.

Instead of logging alerts, the system connects directly to the operating system firewall using Linux iptables. When an anomaly is detected and confirmed, the script automatically executes a temporary DROP rule for the suspected IP.

The platform is evaluated in a local test environment using hping3 and custom scripts to simulate attacks while running on normal background activities. The tests showed that the training on the local baseline eliminated excessive false positives seen when using generic datasets. The autoencoder successfully detected the high reconstruction errors and triggered a firewall drop rule within milliseconds, protecting the host while allowing legitimate traffic to continue uninterrupted

Authors: Rares Muntean
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

16:00 - 16:15

Cristina Dogaru

Smarter Ethereum Wallets and Phishing Risks with EIP 7702

On-line
Digital wallets are often the first point of contact between users and blockchain applications, yet everyday tasks can involve unfamiliar steps and confusing approval requests. This presentation examines EIP-7702, introduced as part of Ethereum's Pectra upgrade, as a way to make existing accounts more flexible and easier to use. It focuses on what these changes mean for ordinary users and on the security questions that arise when a familiar wallet gains additional capabilities.
EIP-7702 allows an ordinary Ethereum account to delegate execution to smart contract code. Depending on the wallet implementation, this can support features such as completing several actions in one transaction or having an application pay transaction fees on behalf of a user. The presentation will explain these possibilities through familiar payment situations, with an emphasis on their practical value rather than the underlying technical design. An illustrative scenario involving two payments will show how a simpler interaction can reduce the number of separate steps a user must complete.
Greater flexibility also makes authorisation decisions more consequential. A request presented as a routine wallet improvement may instead direct the user to malicious functionality. The presentation will examine how phishing can exploit unclear messages, misleading websites and misplaced trust in an apparently legitimate application. It will explain why approving wallet functionality deserves careful attention and how the consequences may extend beyond the single action a user believes they are confirming.
The project will use an accessible overview and a planned demonstration in a controlled environment with test funds. The demonstration will illustrate legitimate use, while the discussion will consider how a deceptive authorisation request could put assets at risk. It will also address the role of understandable wallet prompts, checks on the requested functionality and awareness of the source of a request. These measures will be discussed as practical precautions rather than guarantees of safety.
The intended contribution is an explanation of the relationship between convenience and informed consent in Ethereum wallets. By the end of the presentation, participants should understand what EIP-7702 enables, recognise the significance of delegation requests and identify questions to ask before approving them. The presentation is intended for students and other participants seeking a general understanding of the topic without requiring experience in smart contract development.

Authors: Cristina Dogaru
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

16:15 - 16:30

George Alexandru Dumitru

Web application with authentication and Firebase backend

On-line
This project focuses on the design and development of a secure, cloud-based web application using Firebase as its backend infrastructure. The application aims to provide users with a reliable platform for recording and managing personal activities through an activity log, while demonstrating the integration of modern authentication, database management, serverless computing, and cloud hosting technologies. Firebase Authentication is implemented to support secure user sign-in through email and password credentials, as well as Google accounts. Cloud Firestore is used as the primary NoSQL database for storing and managing user-specific information in real time. To ensure data privacy and security, Firestore Security Rules are configured to restrict users' access exclusively to their own data, preventing unauthorized reading or modification of information. Furthermore, a Cloud Function is developed to automatically execute when a new document is created in Firestore, demonstrating the use of event-driven serverless architecture for automated backend operations. The application is deployed using Firebase Hosting, while the Firebase Emulator Suite is used to simulate Firebase services, test authentication, validate database operations, verify security rules, and evaluate Cloud Function behavior in a controlled local environment before deployment. By integrating these services, the project demonstrates how Firebase can simplify full-stack web application development while providing scalability, security, and reduced backend infrastructure management. The resulting application highlights the practical benefits of serverless cloud technologies and provides an example of how secure, maintainable, and scalable web applications can be developed using the Firebase ecosystem.

Authors: George Alexandru Dumitru
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

16:30 - 16:45

Diana Ioana Dospinescu

Multimodal AI-Assisted Speech Practice Using Large Language Models and Retrieval-Augmented Generation

On-line
Acquired communication impairments following severe traumatic injuries can significantly affect an individual's independence, social interaction, and quality of life. Although speech-language therapy plays an essential role in rehabilitation, access to frequent and personalized practice may be limited by geographical, financial, and time constraints. This study presents an LLM-based multimodal assistive application designed to support gradual speech practice through interactive, voice-enabled exercises. The system combines a speech-capable Large Language Model (LLM) with a visual interface that displays images alongside corresponding words, facilitating the association between visual representations and spoken language. During each exercise, the LLM provides an audio pronunciation of the target word, after which the user is prompted to repeat it. The system evaluates the pronunciation attempt and provides feedback, while progressively adapting exercise difficulty according to the user's performance. For example, successful pronunciation of five predefined easy words enables progression towards more challenging vocabulary. Additionally, a Retrieval-Augmented Generation (RAG) component integrates children's stories and fairy tales, allowing the application to read familiar narratives aloud and provide contextual exposure to spoken language. By combining interactive pronunciation exercises, visual assistance, and narrated storytelling, the proposed application aims to provide accessible and personalized opportunities for communication practice, potentially supporting continued engagement and independent practice outside conventional therapy sessions.

Authors: Diana Ioana Dospinescu
Affiliation: National Institute for Research and Development for Optoelectronics-INOE 2000

16:45 - 17:00

Marius-Laurentiu Popescu

Behavioral Equivalence as a Reconstruction Target: Recovering a Lost Online Game Server Stack from Its Binaries

On-line
Source code for long-lived online services is often lost while the compiled binaries survive. We present a methodology for the functional reconstruction of a complete commercial multiplayer online game backend (ten cooperating Windows x64 server executables built in 2020) into portable, maintainable C++20 source. Instead of aiming for byte-identical machine code, we define equivalence by observable behavior, in priority order: wire protocol (opcodes, field layouts, message ordering), data formats (resource tables, configuration, Lua scripts), database effects, and game logic. Reconstruction follows a strict evidence hierarchy: verified binary identity (PDB GUID/age matching), direct machine-code and runtime evidence, debug-symbol type and line information, cross-reference analysis, and only then explicit, confidence-rated inference. Every recovered function is tagged by its evidence source, and post-build binary patches are detected and preserved. Platform-specific mechanisms (asynchronous networking, Windows services, crash-dump handling) are isolated behind a thin abstraction layer, so one source tree builds and runs on both Windows and Linux. Validation relies on interoperation rather than binary comparison: rebuilt components are swapped one at a time into a deployment of original peers, and captured traffic is replayed and compared byte by byte. In the completed database agent, 883 functions were verified, with replayed responses matching the original on both platforms. This shows that behavioral equivalence at protocol and data boundaries is a practical, verifiable target for recovering lost server software, and that the methodology generalizes to other legacy networked systems.

Authors: Marius-Laurentiu Popescu
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

17:00 - 17:15

Mihail Daniel Bentoiu

Elastic Cloud-Native Architectures: Containerization and Event-Driven Autoscaling of Microservices in Azure Container Apps via KEDA

On-line
Modern cloud-native computing faces a persistent architectural tension: static over-provisioning incurs severe financial waste during idle periods, whereas pure scale-to-zero serverless platforms penalize transient traffic surges with disruptive cold-start latencies. This study investigates the operational elasticity and quality-of-service (QoS) preservation of containerized microservices orchestrated via Azure Container Apps and Kubernetes Event-Driven Autoscaling (KEDA). We implement an asynchronous, hardened ASGI RESTful service using Python and a non-root multi-stage Docker build, distributed through Azure Container Registry and exposed via an Envoy Layer-7 ingress. By configuring a reactive horizontal scaling policy triggered by real-time HTTP connection concurrency rather than delayed resource utilization metrics, the system balances immediate responsiveness with strict budgetary bounds. Empirical burst evaluation using 150 concurrent asynchronous requests demonstrates deterministic horizontal scale-out from a baseline warm replica without gateway timeouts or dropped connections, achieving a 100% request success rate. The findings delineate the trade-offs between warm-standby overhead and latency degradation, providing concrete architectural principles for deploying cost-effective, resilient serverless containers capable of autonomous load absorption in volatile production environments.

Authors: Mihail Daniel Bentoiu
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

17:15 - 17:30

Eduard Bothazi

Design and Implementation of a Cloud-Based Conversational AI System Using Retrieval-Augmented Generation and Semantic Search

On-line
The rapid evolution of artificial intelligence and large language models (LLMs) has significantly advanced the development of intelligent conversational systems. However, conventional language models continue to face challenges related to factual accuracy, hallucinated information, limited access to domain-specific knowledge, and the generation of responses without verifiable sources. Retrieval-Augmented Generation (RAG) offers a promising approach to addressing these limitations by combining semantic information retrieval with generative artificial intelligence, enabling responses grounded in external knowledge sources rather than relying exclusively on the model's internal training data.
This paper presents the design and implementation of a cloud-based conversational AI system integrating Retrieval-Augmented Generation and semantic search to improve the relevance, reliability, and traceability of responses generated from user-provided PDF documents. The proposed solution follows a modular architecture built on Google Cloud technologies, integrating Google Cloud Storage, Vertex AI, Gemini language models, and vector search services.
The system implements a document-processing pipeline that extracts textual information from PDF files, divides the content into overlapping segments, and generates high-dimensional vector embeddings representing semantic relationships. These embeddings are stored alongside document metadata in a vector database, enabling similarity-based information retrieval. When a user submits a question, the system identifies relevant document fragments and provides them as contextual information to the Gemini language model, supporting the generation of contextually relevant answers.
To mitigate hallucinations and reduce the risk of fabricated source references, the architecture incorporates document-grounded response generation, citation identifier validation, and mechanisms for declining to answer questions when sufficient supporting information is unavailable.
A Streamlit-based conversational interface enables users to interact with the system and inspect the document fragments associated with generated responses, improving transparency and verifiability.
Preliminary experiments were conducted using PDF documents, processed into text segments and represented as vector embeddings. Functional testing demonstrated successful semantic retrieval, document-grounded answer generation with identifiable source references, and appropriate refusal to answer a question unrelated to the available document. Furthermore, automated backend unit tests were successfully executed, validating the core embedding, retrieval, generation, and RAG orchestration components.
The proposed modular and cloud-oriented architecture provides a foundation for future scalability, extensibility, and integration with larger document collections. Although systematic evaluation of hallucination reduction and large-scale performance remains future work, the preliminary results demonstrate the feasibility of integrating cloud-based RAG and semantic search into document-oriented conversational AI applications.
The study highlights the importance of architectural modularity, response transparency, reliable information retrieval, and the potential of cloud-based RAG systems in the development of more trustworthy artificial intelligence systems.

Authors: Eduard Bothazi
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

17:30 - 17:45

Nicoleta Iacob

Autonomous Databases: Emerging Paradigms in Artificial Intelligence

On-line
The rapid evolution of artificial intelligence is transforming traditional database management systems into autonomous, adaptive, and intelligent architectures. This paper explores emerging AI-driven paradigms in autonomous databases, focusing on self-optimization, automated query processing, adaptive resource management, and anomaly detection. The research examines the potential of machine learning and generative artificial intelligence to enhance database performance, scalability, and reliability while minimizing human intervention. Furthermore, it identifies key challenges and future research directions for developing efficient, secure, and resilient autonomous database systems.
Keywords: autonomous databases, artificial intelligence, machine learning, database optimization, adaptive systems, generative AI.

Authors: Nicoleta Iacob, Luminita Defta
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

17:45 - 18:00

Luminita Defta

Who Is Responsible for Code Written by AI Agents? Risks and Verification Methods in Software Development

On-line
More and more development teams give AI agents entire tasks: editing files, running tests and opening pull requests. This paper starts from a practical question: how do we make sure this code is correct and secure, and who is responsible when it isn't? We look at the most common problems in generated code, such as subtle logic errors and vulnerable dependencies. We also discuss why traditional code review is no longer enough. Finally, we propose a few working rules for teams: automated testing, mandatory human review for critical areas and a clear record of the changes made by agents.

Authors: Luminita Defta, Nicoleta Iacob
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

18:00 - 18:15

Mihaela Poienariu

Beyond perimeter defense in converged OT–IT Systems: invariant subspaces and active watermarking

On-line
While network segmentation and Industrial DMZs restrict unauthorized perimeter traversal between IT and OT domains, they offer no protection once an adversary gains access to unauthenticated fieldbus networks. Under strict process-latency constraints where cryptography is impractical, anomaly detection relies on Kalman-filter residual evaluation. This note presents an analytical formulation of stealthy false data injection within discrete-time linear systems. We show how an attacker with model knowledge exploits the invariant null space of the residual generator, rendering physical deviations undetectable to classical χ^2 tests. Finally, we formulate the algebraic mechanism by which physical dynamic watermarking restores detectability, establishing the inherent trade-off between attack detection and steady-state control variance.

Authors: Mihaela Poienariu
Affiliation: Faculty of Engineering and Computer Science, SPIRU HARET University

18:15 - 18:30

Laurentiu-Bogdan Asalomia

Integrating Power Quality Monitoring and Machine Learning for Predictive Maintenance of Maritime Navigation Equipment

On-line
Electronic navigation equipment depends on both stable electrical supply and reliable exchange of operational data. Traditional diagnostics often examine power quality disturbances and navigation-system alarms separately, making cross-domain fault patterns difficult to interpret. This paper proposes a monitoring framework that combines voltage, frequency and total harmonic distortion with bridge-side alert and communication indicators. Two unsupervised Isolation Forest detectors, trained on normal operating profiles, are compared with a conventional voltage threshold in a reproducible synthetic benchmark. The experiment contains 12,000 normal training windows, 4,000 normal validation windows and 9,000 test windows, including 1,057 synthetically injected anomalous windows. The fused detector achieves an F1 score of 0.586 and recall of 0.445, compared with 0.394 and 0.261 for the power-quality-only detector. The results show the value of cross-domain features in this controlled setting, while also revealing that fusion does not improve detection of every disturbance category. Because the data are generated rather than recorded onboard, the findings do not establish shipboard effectiveness. The paper concludes with a validation protocol and practical safeguards for operator-facing predictive-maintenance support.

Authors: Laurentiu-Bogdan Asalomia
Affiliation: Northern Marine, Scotland

18:30 - 18:45

Popescu Adi-George

The Dencun Upgrade: EIP-4844 and Blobs

On-line
High transaction fees and network congestion have historically challenged Ethereum's ability to scale for mass adoption. The Dencun upgrade addresses this bottleneck through Ethereum Improvement Proposal 4844 (EIP-4844), also known as proto-danksharding. This presentation explores the technical evolution of Layer 2 rollups, such as Arbitrum, Optimism, and Base, and which will examines how they transitioned from expensive, permanent calldata storage to temporary Binary Large Objects ("blobs"). Unlike traditional transaction data, blobs are pruned after approximately 18 days, drastically reducing storage overhead on Ethereum nodes and creating an independent fee market. As a result, data availability costs have dropped significantly, lowering end-user Layer 2 transaction fees to fractions of a cent. Alongside a conceptual breakdown, this session includes a live Python demonstration querying mainnet blocks via JSON-RPC to inspect real-time blob gas metrics (blobGasUsed and excessBlobGas). Attendees will leave with a clear understanding of how EIP-4844 lays the foundational groundwork for Ethereum's long-term, rollup-centric roadmap.

Authors: Popescu George-Adi
Affiliation: Faculty of Engineering and Computer Science

18:45 - 19:00

Xu Xianda

Automated Invoice Data Extraction Using Azure AI Document Intelligence

On-line
Manual processing of financial and accounting documents remains a time-consuming and error-prone bottleneck for organizations handling large volumes of heterogeneous invoices. This paper presents an automated, cloud-native solution for extracting structured key-value fields and line-item data from scanned or digital PDF and image invoices using Microsoft Azure AI Document Intelligence. Leveraging a Platform-as-a-Service (PaaS) and serverless architecture, the proposed system ingests documents securely via HTTPS into Azure Blob Storage, generates temporary Shared Access Signature (SAS) tokens for controlled access, and processes invoices through the pre-trained prebuilt-invoice deep learning model. Extracted semantic entities—including vendor name, tax identification number, invoice identifier, issue date, line items, VAT amount, and total payable—are validated against confidence score thresholds and persisted as normalized JSON documents in a serverless Azure Cosmos DB NoSQL database for downstream accounting export and reporting.

Authors: Xu Xianda
Affiliation: Faculty of Engineering and Computer Science

19:00 - 19:15

Ana Maria Ramona Stancu

Fuzzy Sets in Machine Learning and Data Mining

On-line
In recent years, models for machine learning and knowledge extraction have garnered attention within the research community, particularly in the field of fuzzy sets. This paper reviews several algorithms from fuzzy set theory that contribute to machine learning and data mining-as applied to datasets-and concludes with suggestions for future development directions.

Authors: Ana Maria Ramona Stancu, Drinceanu Maria Alexandra
Affiliation: Faculty of Engineering and Computer Science

Tudor Vianu National College of Computer Science, Bucharest


19:15 - 19:30

Corina Elena Vint

From Proof to Algorithm: Mathematical Reasoning in High School Computer Science

On-line
This paper proposes an interdisciplinary approach in which high school students use mathematical reasoning to design and justify algorithms. The learning activities focus on two examples, the Euclidean algorithm and binary search, and guide students from the underlying mathematical property to its implementation in code. For the Euclidean algorithm, students explain why the greatest common divisor is preserved when a pair of numbers is replaced by the divisor and the remainder. For binary search, they justify why half of the search interval in a sorted sequence can be discarded at each step. Students then test their implementations on typical inputs and edge cases. The approach aims to develop students’ ability to explain why an algorithm works and identify the conditions under which it produces the correct result.

Authors: Corina Elena Vint, Alina Gabriela Boca
Affiliation: Tudor Vianu National College of Computer Science, Bucharest

19:30 - 19:45

Corina Elena Vint

From Probability to Simulation: Exploring Random Phenomena Through High School Programming

On-line
This paper proposes a series of activities in which high school students investigate the relationship between theoretical probability and relative frequency obtained through simulation. Starting with a question such as “How often do two dice add up to seven?”, students identify the possible outcomes, calculate the probability, and write a Python program to repeat the experiment. They compare results across different numbers of trials and repeat the simulations to observe variation between runs. The discussion focuses on how relative frequencies tend to approach theoretical probabilities as the number of trials increases, even though the difference can fluctuate from one run to another. The activities connect mathematical modelling with programming and data interpretation.

Authors: Corina Elena Vint, Alina Gabriela Boca
Affiliation: Tudor Vianu National College of Computer Science, Bucharest

19:45 - 20:00

Laura Ungureanu

Intelligent Agents in Schools: Assisting Librarians and Administrative Staff in Their Daily Work

On-line
Anyone who has worked in a school library or administrative office knows how frequently the same questions arise. "Do you have this book?" "Where can I find the school regulations?" "When can I pick up my certificate?" Each individual request may seem minor; however, taken together, they consume a significant portion of the day, leaving little time for tasks that truly matter. This article addresses this challenge by proposing a practical solution: a team of five artificial intelligence agents that work together, much like a small, collaborative staff unit.

The first agent acts as a coordinator—listening to the request, identifying the user's needs, and delegating tasks to the other agents. The second agent handles the search process, scouring the library catalog, archives, and administrative records to locate the required information. The third agent verifies that the retrieved information is accurate and up-to-date, while also ensuring that the requester is authorized to access it—a critical step when personal data is involved. The fourth agent drafts a response tailored to the specific user, whether they are a fifth-grade student, a teacher, or a parent. The fifth agent maintains a record of all requests and their resolutions, allowing the system to learn from experience and become increasingly useful over time.

We present several real-world scenarios from a school environment and discuss the benefits: faster responses, fewer errors, and the ability to access information even outside of regular working hours. However, we do not overlook the difficulties either. Such a system only works if the data is properly digitized. It requires human oversight and raises questions regarding privacy. We believe that these agents do not replace the librarian or the secretary; rather, they relieve them of repetitive tasks, leaving them more time for the people they serve.

Authors: Laura Ungureanu
Affiliation: Tudor Vianu National College of Computer Science, Bucharest

20:00 - 20:15

Grigore Albeanu
Marius Iulian Mihailescu

CAMAI 2026 Closing and Remarks

On-line
Closing session of CAMAI 2026: a review of the two days of talks, best-paper and best-presentation acknowledgements, and an outline of the next edition of the conference.

Submission & Registration

Scientific researchers from the universities, research centers, industry and education institutions are invited to send original research contributions to be considered for review and presentation in CAMAI 2026 sections.

Submission of paper abstracts is equivalent with pre-registration.

Important dates for CAMAI 2026
Abstract submission closes30 October 2026
Notification of acceptance6 November 2026
Final version due13 November 2026
Conference opens20 November 2026

Abstracts will be sent as PDF files and will include the following information:

  • title
  • name and affiliation for all authors
  • e-mail for corresponding author
  • abstract
  • keywords
  • AMS or ACM classification

Abstracts will be in English. Abstract length must be between 1 and 2 pages (A4 size).

Abstracts and full papers go through EasyChair, which tracks your submission and the reviewers' response in one place. Everything you need is in the Submit a Paper section below.

For questions the platform cannot answer, write to secretariat_misn@spiruharet.ro

Final registration is confirmed by e-mail once your paper is accepted.


Event Starts In:

Submit a Paper

Format your paper with the CAMAI 2026 template, export it to PDF, then upload it to EasyChair. Papers that do not follow the template are returned before review.

EasyChair is where every submission is handled

Create a free EasyChair account if you do not have one, select CAMAI 2026, and upload your PDF. You can revise it as often as you like until the deadline, and the reviewers' response arrives in the same place.

Go to EasyChair
Word template

Page size, margins, heading levels and the reference style are already set. Type into it rather than restyling your own document.

LaTeX template

The same layout as a LaTeX source file: 170 by 235 mm page, the Annals heading block, the abstract and keywords structure, and a worked reference list.

Before you upload

Check that the title, every author name and affiliation, the corresponding e-mail, the abstract, keywords and an AMS or ACM classification are all present, and that the file is a PDF.

Trouble with the upload? Send the paper to secretariat_misn@spiruharet.ro and the secretariat will register it for you.

Editorial Committee

Grigore Albeanu, Dr.
Professor
Dean of Engieering and Computer Science Faculty
Iordan Duda, Dr.
Professor
Engieering and Computer Science Faculty

Scientific Committee

  • Prof. Dr. Vasile STAICU, University of Aveiro, Portugal
  • Prof. Dr. Gregory MAKRIDES, University of Cyprus, Cyprus
  • Prof. Dr. Anuradha MAHASINGHE, University of Colombo, Sri Lanka
  • Sci. Res. I Dr. Veturia CHIROIU, Institute of Solid Mechanics of The Romanian Academy, Romania
  • Sci. Res. I Dr. Ligia MUNTEAN, Institute of Solid Mechanics of The Romanian Academy, Romania
  • Prof. Dr. Vasile PREDA, University of Bucharest, Romania
  • Prof. Dr. Ion CHITESCU, University of Bucharest, Romania
  • Prof. Dr. Tudor BALANESCU, University of Pitesti, Romania
  • Prof. Dr. Horia GEORGESCU, University of Bucharest, Romania
  • Assoc. Prof. Dr. Rodica IOAN, Spiru Haret University, Bucharest, Romania
  • Assoc. Prof. Dr. Marius Iulian MIHAILESCU, Spiru Haret University, Bucharest, Romania

Archive

Venue

How To Get Here

Facilities & Accommodations

Search on Booking Search on AirBnb