黑料不打烊

艩is bakalauro baigiamasis darbas pateikia lyginam膮j寞 naudojamumo vertinim膮 dviej懦 pla膷iai naudojam懦 priekin臈s s膮sajos vartotojo s膮sajos karkas懦 鈥� Material UI ir Bootstrap. Tyrimas nagrin臈ja, kaip skirtingi vartotojo s膮sajos karkasai veikia suvokiam膮 naudojamum膮, s膮veikos patirt寞 ir vartotoj懦 preferencijas, kai jie taikomi interaktyviems 啪iniatinklio prototipams, sukurtiems kontroliuojamomis s膮lygomis. Pirmasis skyrius pristato pagrindines vartotojo s膮sajos dizaino ir 啪mogaus鈥搆ompiuterio s膮veikos s膮vokas, 寞skaitant naudojamumo principus, prieinamum膮, kognityvin臋 apkrov膮, vizualin臋 hierarchij膮 ir komponent懦 nuoseklum膮. Taip pat ap啪velgiami susij臋 tyrimai apie vartotojo s膮sajos karkasus ir naudojamumo vertinimo metodus. Antrasis skyrius apra拧o tyrimo metodologij膮, 寞skaitant naudojamumo klausimyn懦 sudarym膮, dalyvi懦 atrank膮 ir duomen懦 rinkimo proced奴ras. Tre膷iasis skyrius pristato dviej懦 啪iniatinklio prototip懦 k奴rim膮 naudojant Material UI ir Bootstrap karkasus. Abu prototipai buvo sukurti su identi拧ka strukt奴ra, turiniu ir navigacijos srautu, siekiant u啪tikrinti kontroliuojam膮 palyginim膮, orientuot膮 寞 dizaino 寞gyvendinim膮. Ketvirtasis skyrius pateikia 55 dalyvi懦 atlikto naudojamumo vertinimo rezultatus. Analiz臈 apima apra拧omuosius ir inferencinius statistinius metodus. Penktasis skyrius aptaria rezultatus naudojamumo teorijos kontekste, pabr臈啪damas, kaip s膮sajos dizaino veiksniai, tokie kaip vizualin臈 hierarchija, tarp懦 i拧d臈stymas ir komponent懦 nuoseklumas, veikia vartotojo patirt寞 ir suvokim膮. Baigiamojo darbo apimtis: 52 puslapiai be pried懦, 14 paveiksl懦, 5 lentel臈s, 20 literat奴ros 拧altini懦.

This bachelor thesis presents a comparative usability evaluation of two widely used front end user interface frameworks, Material UI and Bootstrap. The study investigates how different UI frameworks influence perceived usability, interaction experience, and user preferences when applied to interactive web prototypes developed under controlled conditions. The first chapter introduces fundamental concepts of user interface design and human-computer interaction, including usability principles, accessibility, cognitive load, visual hierarchy, and component consistency. It also reviews related research on UI frameworks and usability evaluation methods. The second chapter describes the research methodology, including the design of usability questionnaires, participant selection, and data collection procedures. The third chapter presents the development of two web-based prototypes using Material UI and Bootstrap frameworks. Both prototypes were designed with identical structure, content, and navigation flow to ensure a controlled comparison focused on design implementation. The fourth chapter presents the results of the usability evaluation conducted with 55 participants. The analysis includes descriptive and inferential statistical methods. The fifth chapter discusses the results in relation to usability theory, highlighting how interface design factors such as visual hierarchy, spacing, and component consistency influence user experience and perception. Thesis volume: 52 pages excluding annexes, 14 figures, 5 tables, 20 bibliography entries.

艩iame bakalauro baigiamajame darbe bus nagrin臈jamas kliento UAB 鈥災甿on臈l臈鈥� fizini懦 serveri懦 konsolidavimo ir atk奴rimo po nelaimi懦 sprendimo k奴rimas ir testavimas pasitelkiant debes懦 kompiuterijos platformas. Darbo tikslas 鈥� sukurti ir pagr寞sti architekt奴r膮, kuri leid啪ia konsoliduoti kliento fizinius serverius ir apsispr臋sti d臈l galimyb臈s laikyti savo virtualius serverius debesijos paslaugas teikian膷ioje bendrov臈je. Pirmiausia analizuojama esama infrastrukt奴ra, jos apkrovimas, naujumas, verslo t臋stinumo reikalavimai. Remiantis gauta informacija projektuojama hibridin臈 architekt奴ra (virtualizacija, saugyklos, tinklai, duomen懦 sauga), argumentuojamas debes懦 paslaug懦 ir technologij懦 pasirinkimas, bei parengiamas migracijos planas ir rizikos valdymo matrica. Praktin臈je 拧io darbo dalyje 寞gyvendinamas bandomasis atk奴rimo sprendimas debesyje, bei duomen懦 replikavimas ir 鈥瀎ailover/failback鈥� testai. Atlikti bandomieji atstatymai leid啪ia 寞vertinti sprendimo efektyvum膮 pagal pasiektus RTO/RPO, bei 寞vertinti sistem懦 na拧umo poky膷ius ir bendr膮j寞 TCO. Gauti rezultatai leid啪a pateikti praktines rekomendacijas diegimui ir naudojimui, bei nurodo optimizavimo gaires. Darbo vert臈 鈥� ai拧kus kelias nuo esamos pad臈ties analiz臈s iki patikrinto atsakingo u啪 kliento 鈥災甿on臈l臈鈥� DR sprendimo, bei ekonomi拧kumo palyginimo, kur寞 galima taikyti pana拧iuose projektuose. Darb膮 sudaro 9 dalys: 寞vadas, esamos infrastrukt奴ros analiz臈, galim懦 sprendim懦 ap啪valga ir palyginimas, architekt奴ros projektavimas debes懦 paslaugose, migracijos ir konsolidavimo planavimas, bandomasis disaster recovery sprendimas, i拧vados ir literat奴ros s膮ra拧as.

This bachelor's thesis examines the development and testing of a physical server consolidation and disaster recovery solution for the client UAB "漠mon臈l臈" using cloud computing platforms. The objective of this work is to design and justify an architecture that enables the consolidation of the client's physical servers and to determine the feasibility of hosting their virtual servers with a cloud service provider. First, the existing infrastructure, its load, age, and business continuity requirements are analyzed. Based on the gathered information, a hybrid architecture (virtualization, storage, networks, data security) is designed, the selection of cloud services and technologies is justified, and a migration plan is prepared alongside a risk management matrix. In the practical part of this thesis, a pilot disaster recovery solution is implemented in the cloud, along with data replication and failover/failback testing. The trial recoveries make it possible to evaluate the effectiveness of the solution based on the achieved RTO/RPO, as well as to assess changes in system performance and the overall TCO. The obtained results provide practical recommendations for implementation and operation, and outline guidelines for further optimization. The value of this work lies in providing a clear path from the analysis of the current situation to a validated disaster recovery (DR) solution for the client "漠mon臈l臈," including an economic comparison that can be applied to similar projects.

Baigiamajame bakalauro darbe sukurta kompiuterini懦 tinkl懦 atak懦 klasifikavimo sistema, gebanti automati拧kai atpa啪inti atak懦 tipus pagal tinklo srauto po啪ymius. Darbo tikslas: sukurti sistem膮, kuri naudojasi ma拧ininio mokymosi modeliu kompiuteri懦 tinklo atak懦 klasifikavimui 寞mon臈je 鈥濨altijos pa啪angi懦 technologij懦 institutas鈥�. Analitin臈je dalyje ap啪velgta tinklo srauto paket懦 strukt奴ra, nustatyti pagrindiniai atak懦 tipai ir i拧nagrin臈ti populiar奴s ma拧ininio mokymosi modeliai. Projektin臈je dalyje pateikiami sistemos reikalavimai, parinktos technologijos, pasirinktas atsitiktini懦 mi拧k懦 ma拧ininio mokymo modelis ir apra拧yta sistemos architekt奴ra. Realizacijos dalyje apra拧ytas sistemos komponent懦 veikimas: tinklo srauto fiksavimas, po啪ymi懦 i拧vedimas, ma拧ininio mokymo modelio mokymas ir vartotojo s膮saja. Testavimo dalyje pateikti rezultatai, gauti vykdant a拧tuonias skirtingas kompiuterines atakas izoliuotoje virtualioje aplinkoje naudojan膷ioje 鈥濴inux Ubuntu鈥� operacin臋 sistem膮, kurie patvirtino sistemos funkcionalum膮. Darb膮 sudaro septynios dalys: 寞vadas, technin臈 ap啪valga, ma拧ininio mokymosi sprendim懦 ap啪valga, projektavimas, realizacija, testavimas ir i拧vados. Darbo apimtis: 62 p. teksto be pried懦, 30 pav., 8 lent., 38 bibliografiniai 拧altiniai. Atskirai pridedami darbo priedai.

This bachelor鈥檚 thesis presents a computer network attack classification system capable of automatically identifying attack types based on network traffic features. The aim of the thesis is to develop a decision support tool that helps monitor network traffic in real time and identify ongoing attacks at the company 鈥淏altic Institute of Advanced Technology鈥�. The analytical part reviews the structure of network traffic packets, identifies the main types of attacks, and examines popular machine learning models. The design part presents the system requirements, the selected technologies, the chosen 鈥淩andom Forest鈥� model, and the system architecture. The implementation part describes the operation of the system components: network traffic capture, feature extraction, model training, and the user interface. The testing part presents the results obtained by executing eight different computer attacks in an isolated virtual environment running the 鈥淟inux Ubuntu鈥� operating system, which confirmed the suitability of the system. The thesis consists of seven parts: introduction, technical overview, overview of machine learning solutions, design, implementation, testing, and conclusions. Thesis volume: 62 pages of text, 30 figures, 8 tables, 38 bibliographic sources. Appendices are attached separately.

Baigiamajame bakalauro darbe analizuojamas kei膷iamo mastelio debes懦 infrastrukt奴ros projektavimas ir diegimas, taikant infrastrukt奴ros kaip kodo (Infrastructure as Code) metod膮 bei CI/CD automatizavim膮. Darbe nagrin臈jami pagrindiniai debes懦 kompiuterijos, 鈥濪evOps鈥�, infrastrukt奴ros kaip kodo, CI/CD ir konfig奴racijos valdymo principai. Debes懦 paslaug懦 teik臈j懦, infrastrukt奴ros kaip kodo 寞ranki懦, CI/CD platform懦 ir konfig奴racijos valdymo 寞ranki懦 palyginimas atliekamas pagal pasirinktus vertinimo kriterijus. Remiantis palyginimo rezultatais, praktiniam 寞gyvendinimui pasirinkti AWS, 鈥濼erraform鈥�, 鈥濲enkins鈥� ir 鈥濩hef鈥�. Pasirinktos technologijos taikomos projektuojant ir 寞gyvendinant kei膷iamo mastelio debes懦 infrastrukt奴r膮. 鈥濼erraform鈥� naudojamas infrastrukt奴ros resursams AWS aplinkoje kurti, 鈥濩hef鈥� 鈥� sukurt懦 instancij懦 konfig奴ravimui, o 鈥濲enkins鈥� koordinuoja automatizuotus darbo procesus. 漠gyvendintas sprendimas vertinamas pagal mastelio keitimo galimybes, atkuriamum膮 ir diegimo efektyvum膮. Atlikus analiz臋, technologij懦 pasirinkim膮 ir praktin寞 寞gyvendinim膮, pateikiamos baigiamojo bakalauro darbo i拧vados. Darb膮 sudaro 7 dalys: esamos situacijos analiz臈, technologij懦 pasirinkimas, sistemos reikalavim懦 specifikacija ir 寞gyvendinimo planas, sistemos architekt奴ra, i拧vados ir literat奴ros 拧altini懦 s膮ra拧as. Darbo apimtis 鈥� 48 p. teksto be pried懦, 6 iliustr, 26 lent., 64 bibliografiniai 拧altiniai. Atskirai pridedami darbo priedai.

In the final bachelor thesis, the design and implementation of scalable cloud infrastructure using Infrastructure as Code approach and CI/CD automation are analyzed. The work inspects the main principles of cloud computing, DevOps, Infrastructure as Code, CI/CD and configuration management. The comparison of cloud providers, Infrastructure as Code tools, CI/CD platforms and configuration management tools is based on selected evaluation criteria. Based on comparison results AWS, Terraform, Jenkins and Chef are selected for the practical implementation. The selected technologies are applied to design and implement scalable cloud infrastructure. Terraform provisions the infrastructure resources in AWS. Chef configures the created instances. Jenkins orchestrates the automation workflows. The implemented solution is assessed according to scalability, reproducibility and deployment efficiency. After analysis, technology selection and practical implementation, the conclusions of the final bachelor thesis are presented. The thesis consists of 7 parts: analysis of current situation, technology selection, system requirements specification and implementation plan, system architecture, conclusions, list of references. The thesis consists of 48 pages without appendices, 6 figures, 26 tables and 64 bibliographic sources. Appendices are provided separately.

Baigiamajame magistro darbe nagrin臈jami studij懦 nutraukimo prognozavimo metodai, naudojant 黑料不打烊 student懦 duomenis, kuriuos sudaro 1761 寞ra拧ai ir 33 po啪ymiai. Darbo tikslas buvo sukurti i拧vestinius elgsenos po啪ymius, 寞vertinti j懦 taikymo naud膮 studij懦 nutraukimo prognozavime ir nustatyti optimal懦 ma拧ininio mokymosi modelio, duomen懦 rinkinio bei hiperparametr懦 derin寞. Tyrime analizuoti logistin臈s regresijos ir atsitiktinio mi拧ko modeliai, taip pat vertintos duomen懦 tvarkymo strategijos ir po啪ymi懦 grupi懦 寞taka prognozei. Nustatyta, kad i拧vestini懦 po啪ymi懦 寞traukimas modeli懦 veikimo nepagerino. Geriausi膮 rezultat膮 pasiek臈 logistin臈s regresijos modelis, kuriame taikytas 鈥瀕iblinear鈥� optimizavimo metodas, 0,4 klasifikavimo slenkstis ir 0,1 reguliarizacijos stiprumas, pasiektas 0,82 tikslumas. Taip pat nustatyta, kad modelis geba stabiliai identifikuoti studij懦 nutraukimo rizik膮 ankstyvosiomis studij懦 savait臈mis. Darb膮 sudaro 寞vadas, studij懦 nutraukimo prognozavimo mokslin臈je literat奴roje skyrius, tyrimo metodologijos skyrius, tyrimo 寞gyvendinimo ir rezultat懦 skyrius, i拧vados, naudotos literat奴ros s膮ra拧as. Darbo apimtis 鈥� 75 psl., 6 pav., 17 lent., 30 literat奴ros 拧altini懦.

This master鈥檚 thesis investigates student dropout prediction methods using 黑料不打烊 student data consisting of 1,761 records and 33 features. The aim of the study was to develop derived behavioral features, evaluate their benefit for student dropout prediction, and determine the optimal combination of machine learning model, dataset, and hyperparameters. The research analyzed logistic regression and random forest models, as well as data processing strategies and the impact of feature groups on prediction performance. The results showed that the inclusion of derived features did not improve model performance. The best result was achieved by the logistic regression model using the 鈥渓iblinear鈥� optimization method, a classification threshold of 0.4 and a regularization strength of 0.1, reaching an accuracy of 0.82. It was also found that the model is capable of consistently identifying dropout risk during the early weeks of study. The thesis consists of an introduction, a chapter reviewing student dropout prediction in scientific literature, a research methodology chapter, a chapter presenting the implementation and results of the study, conclusions, a list of references. The thesis comprises 75 pages, 6 figures, 17 tables, 30 references.

Baigiamajame magistro darbe nagrin臈jamas automatinis cemento sektoriaus 寞moni懦 akcij懦 kain懦 prognozavimas, remiantis vie拧ai prieinamais rinkos, fundamentiniais, makroekonominiais ir eksporto duomenimis. Darbe analizuojami kain懦 prognozavimo metodai, ap啪velgiami laiko eilu膷i懦, ekonometriniai ir ma拧ininio mokymosi modeliai, taip pat suformuojamas keli懦 拧altini懦 analitinis duomen懦 rinkinys. Tyrimo metu atliekamas duomen懦 valymas, tr奴kstam懦 reik拧mi懦 vertinimas, po啪ymi懦 formavimas ir modeli懦 palyginimas. Kitos savait臈s akcij懦 u啪darymo kainai prognozuoti taikomi baziniai, regresiniai, laiko eilu膷i懦 ir ansambliniai ma拧ininio mokymosi modeliai. Rezultatai parod臈, kad trumpalaik臈je savaitin臈je prognoz臈je geriausiai veik臈 paprasti inerciniai modeliai, ypa膷 paskutin臈s reik拧m臈s principu pagr寞stas modelis. Nustatyta, kad papildomi kontekstiniai duomenys yra naudingi rinkos analizei, ta膷iau vienos savait臈s kainos lygio prognoz臈je j懦 poveik寞 riboja stiprus kainos persistencijos efektas.

This master鈥檚 thesis examines automated stock price prediction for companies in the cement sector, based on publicly available market, fundamental, macroeconomic and export data. The thesis analyses price prediction methods, reviews time series, econometric and machine learning models, and develops a multi-source analytical dataset. The research includes data cleaning, assessment of missing values, feature engineering and model comparison. Baseline, regression, time series and ensemble machine learning models are applied to predict the next week鈥檚 stock closing price. The results showed that simple inertial models performed best in short-term weekly forecasting, especially the model based on the last observed value principle. It was found that additional contextual data are useful for market analysis; however, their impact on one-week price level prediction is limited by the strong persistence effect of stock prices.

艩iame baigiamajame darbe nagrin臈jamas socialini懦 tinkl懦 vaizd懦 analiz臈s metodas, skirtas naudotojo lankom懦 viet懦 nustatymui, taikant dirbtinio intelekto ir kompiuterin臈s regos sprendimus. Pateikiama mokslin臈s literat奴ros ap啪valga, analizuojanti geolokalizacijos metodus, paremtus vaizdais, metaduomenimis ir tekstine informacija, bei aptariami j懦 privalumai ir ribotumai. Pagrindinis darbo tikslas 鈥� sukurti sistemos architekt奴r膮, kuri, pateikus socialinio tinklo naudotojo duomenis, geb臈t懦 nustatyti nuotrauk懦 geografin臋 viet膮, taikant keli懦 lygi懦 metod膮: pirmiausia naudojant EXIF metaduomenis, v臈liau vaizd懦 analiz臋, o galiausiai 鈥� tekstini懦 apra拧懦 apdorojim膮. Modelio mokymui naudojamas 鈥濴arge Dataset of Geotagged Images鈥� duomen懦 rinkinys, kuris yra YFCC100M duomen懦 rinkinio poaibis, o vertinimui 鈥� real奴s naudotojo 鈥濱nstagram鈥� duomenys. Darbo metu taikyti 拧iuolaikiniai metodai, tokie kaip CLIP ir GeoCLIP modeliai, artimiausi懦 kaimyn懦 paie拧ka (FAISS), bei duomen懦 filtravimo ir balansavimo strategijos. Eksperimentai parod臈, kad kombinuotas metodas leid啪ia efektyviai nustatyti vietov臋 net ir nesant tiksli懦 metaduomen懦. Taip pat analizuojami skirting懦 metod懦 tikslumo skirtumai bei j懦 pritaikomumas realiose situacijose. Darbas susideda i拧 寞vado, literat奴ros ap啪valgos, metodikos, eksperimentin臈s dalies ir i拧vad懦. Darbo apimtis 鈥� 94 puslapiai be pried懦, 17 paveiksl懦, 19 lenteli懦, 29 literat奴ros 拧altiniai.

This thesis investigates a method for analyzing social media images to identify locations visited by users, using artificial intelligence and computer vision techniques. A review of scientific literature is presented, analyzing geolocation methods based on visual data, metadata, and textual information, as well as discussing their advantages and limitations. The main objective of this work is to design a system architecture that, given social media user data, can determine the geographical location of images by applying a multi-level approach: first utilizing EXIF metadata, followed by image-based analysis, and finally processing textual descriptions. The model is trained using the Large Dataset of Geotagged Images, which is a subset of the YFCC100M dataset, while evaluation is performed on real user Instagram data. Modern techniques such as CLIP and GeoCLIP models, approximate nearest neighbor search (FAISS), as well as data filtering and balancing strategies are applied in this study. Experimental results demonstrate that the combined approach enables effective location estimation even in the absence of precise metadata. The accuracy of different methods and their applicability in real-world scenarios are also analyzed. The thesis consists of an introduction, literature review, methodology, experimental section, and conclusions. The length of the thesis is 94 pages (excluding appendices), including 17 figures, 19 tables, and 29 references.

Baigiamajame magistro darbe nagrin臈jamos medicinini懦 vaizd懦 duomen懦 stygiaus ir prieinamumo problemos bei j懦 sprendimo galimyb臈s pasitelkiant dirbtinio intelekto metodus. I拧analizuoti pagrindiniai kompiuterin臈s regos modeli懦 taikymo medicinoje principai, medicinini懦 vaizd懦 tipai ir jiems b奴dingi u啪daviniai. Aptartos pagrindin臈s duomen懦 tr奴kumo prie啪astys 鈥� techniniai, laiko bei asmens duomen懦 apsaugos apribojimai. I拧nagrin臈tos tradicin臈s duomen懦 i拧pl臈timo (toliau 鈥� augmentacija) technikos (geometrin臈s ir pikseli懦 lygmens transformacijos) bei modern奴s sintetini懦 duomen懦 generavimo metodai 鈥� generatyviniai prie拧prie拧iniai tinklai (GAN) ir difuzija pagr寞sti modeliai (DDPM). Atlikta naujausi懦 mokslini懦 tyrim懦 ap啪valga skirting懦 medicinini懦 vaizd懦 modalum懦 kontekste. Metodologin臈je dalyje apra拧yti pasirinkti duomen懦 rinkiniai, kompiuterin臈s regos modeliai, augmentacijos strategijos ir tikslumo vertinimo kriterijai. Empirinio tyrimo metu eksperimenti拧kai i拧bandytos ir palygintos skirtingos augmentacijos bei hibridin臈s j懦 kombinavimo strategijos, siekiant nustatyti optimal懦 duomen懦 pasiskirstym膮 ir sugeneruot懦 vaizd懦 proporcijas medicinini懦 vaizd懦 klasifikavimo u啪daviniams. Remiantis gautais rezultatais, pateikiamos i拧vados ir rekomendacijos praktiniam dirbtinio intelekto sprendim懦 pritaikymui sveikatos prie啪i奴roje. Darb膮 sudaro 6 dalys: 寞vadas, literat奴ros analiz臈, metodologija, empirinis tyrimas, i拧vados ir si奴lymai, literat奴ros s膮ra拧as. Darbo apimtis 鈥� 120 p. teksto be pried懦, 41 iliustr., 11 lent., 93 bibliografiniai 拧altiniai.

This Master's thesis addresses the problems of medical image data scarcity and accessibility, exploring potential solutions through artificial intelligence methods. The core principles for applying computer vision models in medicine, along with various medical imaging modalities and their inherent challenges, are analyzed. The primary causes of data deficiency 鈥� including technical, time-related, and personal data protection constraints 鈥� are discussed. Furthermore, traditional data expansion (hereinafter referred to as data augmentation) techniques, such as geometric and pixel-level transformations, are examined alongside modern synthetic data generation methods, namely Generative Adversarial Networks (GANs) and Denoising Diffusion Probabilistic Models (DDPM). A literature review of the latest scientific research across different medical imaging modalities is provided. The methodological section outlines the selected datasets, computer vision models, augmentation strategies, and performance evaluation criteria. During the empirical study, various augmentation techniques and hybrid combination schemes were experimentally tested and compared to determine the optimal data distribution and synthetic image proportions for medical image classification tasks. Based on the obtained results, conclusions and recommendations are presented for the practical application of artificial intelligence solutions within the healthcare sector. The thesis consists of 6 sections: introduction, literature review, methodology, empirical research, conclusions and recommendations, and a list of references. The scope of the thesis is 120 pages of text excluding appendices, and it includes 41 figures, 11 tables, and 93 bibliographical sources.

Bakalauro baigiamajame darbe nagrin臈jami nat奴raliosios kalbos dirbtinio intelekto sistem懦 pa啪eid啪iamumo metodai ir saugumo priemoni懦 veiksmingumas ribojant prieig膮 prie 啪alingo turinio. Teorin臈je dalyje analizuojami did啪i懦j懦 kalbos modeli懦 veikimo principai, taikomi saugumo mechanizmai, j懦 ribotumai bei da啪niausiai naudojami apsaug懦 ap臈jimo (jailbreak) metodai, 寞skaitant vaidmen懦 啪aidimo, semantinio maskavimo, instrukcij懦 hierarchijos manipuliavimo ir automatizuoto optimizavimo technikas. Eksperimentin臈je dalyje atliktas vie拧懦 dirbtinio intelekto sistem懦 (鈥濩hatGPT鈥�, 鈥濭emini鈥� ir 鈥濩laude鈥�) saugumo priemoni懦 vertinimas. Sukurta Python pagrindu veikianti testavimo metodika ir u啪klaus懦 rinkiniai saugumo mechanizm懦 efektyvumui 寞vertinti. Tyrimo metu analizuotas sistem懦 atsparumas skirtingiems pa啪eid啪iamumo metodams, atliktas atsakym懦 klasifikavimo palyginimas tarp platform懦 ir 寞vertintas saugumo vertinimo patikimumas. Remiantis gautais rezultatais, pasi奴lytas rizikos vertinimu gr寞stas metodas 啪alingo turinio aptikimui ir dirbtinio intelekto atsakym懦 saugumo vertinimo patikimumui gerinti. Taip pat naudojant 鈥濻treamlit鈥� bibliotek膮 sukurtas eksperimentinis saugumo filtras. Darbo pabaigoje pateikiamos tyrimo i拧vados, aptariami darbo ribotumai ir numatomos tolesni懦 tyrim懦 kryptys. Darbo apimtis 鈥� 87 p. teksto be pried懦, 33 iliustr., 17 lent., 28 bibliografini懦 拧altini懦. Atskirai pridedami darbo priedai.

The bachelor's final work investigates vulnerability methods of natural language artificial intelligence systems and the effectiveness of security measures in preventing access to harmful content. Theoretical analysis covers the operating principles of large language models, existing security mechanisms, their limitations, and common jailbreak techniques, including role-playing, semantic obfuscation, instruction hierarchy manipulation, and automated optimization methods. An experimental evaluation of security measures implemented in public AI systems (鈥淐hatGPT鈥�, 鈥淕emini鈥�, and 鈥淐laude鈥�) was conducted. A Python-based testing methodology and prompt datasets were developed to assess the effectiveness of security mechanisms. The study analyzed system resilience against different vulnerability methods, compared response classifications across platforms, and evaluated the reliability of security assessments. Based on the results, a risk assessment鈥揵ased approach was proposed to improve harmful content detection and AI response safety evaluation. An experimental security filter was also implemented using the 鈥淪treamlit鈥� library. The work concludes with research findings, limitations, and recommendations for future research. Work volume 鈥� 87 pages of text excluding appendices, 33 figures, 17 tables, 28 bibliographic sources. Appendices are attached separately.

艩iame bakalauro darbe projektuojama, 寞gyvendinama ir vertinama dirbtinio intelekto pagrindu veikianti dokument懦 analiz臈s platforma 鈥濱ntelliDoc", leid啪ianti vartotojams 寞kelti dokumentus ir gauti tiksliais citatomis pagr寞stus atsakymus 寞 nat奴ralios kalbos u啪klausas. Darbe tiriamas informacijos paie拧kos ir generavimo metodas kaip pagrindinis dokument懦 supratimo mechanizmas, integruojant dviej懦 modeli懦 atsakym懦 sintez臋 ir semantin臈s paie拧kos galimybes. Praktinio tyrimo metu sukurta veikianti 啪iniatinklio programa, palaikanti PDF, DOCX ir TXT formatus, bei dokumentuota sistemos architekt奴ra, veikimo 寞vertinimas ir diegimo proced奴ros. Tyrimo tikslas 鈥� sukurti lokaliai veikian膷i膮 dokument懦 analiz臈s priemon臋, teikian膷i膮 tikslius ir patikrinamus atsakymus tiek individualiems vartotojams, tiek organizacijoms, tvarkan膷ioms jautrius duomenis. Sistema naudoja model寞 all-MiniLM-L6-v2 tankiajam vektoriniam paie拧kai, o kry啪minio koduotojo perrangos modelis patikslina kandidat懦 fragmentus prie拧 generuojant atsakym膮. Du nepriklausomi kalbos modeliai vietinis FLAN-T5 ir GROQ debesijos API atsakymus generuoja lygiagre膷iai, o automatinio atrankos mechanizmas parenka labiau pagr寞st膮 atsakym膮. Darb膮 sudaro penki pagrindiniai skyriai: 寞vadas, aktuali懦 technologij懦 analiz臈, sistemos reikalavim懦 specifikacija, sistemos architekt奴ros apra拧ymas, sistemos dokumentacija ir testavimas. Baigiamajame darbe taip pat pateikiamos i拧vados ir literat奴ros s膮ra拧as. Darbo apimtis 鈥� 58 pagrindiniai puslapiai, 11 paveiksl懦, 6 lentel臈s ir 18 bibliografini懦 拧altini懦.

The present bachelor's research designs, implements, and evaluates an AI-powered document intelligence platform named IntelliDoc, which enables users to upload documents and retrieve precise, citation-backed answers using natural language queries. The research investigates retrieval-augmented generation as a core mechanism for document understanding while integrating dual-model answer synthesis and semantic search capabilities. The practical investigation produced a fully functional web application supporting PDF, DOCX, and TXT formats while documenting system architecture, performance evaluation, and deployment procedures. The research aimed to develop a local-first document intelligence tool that delivers accurate and verifiable answers to both individual users and organisations handling sensitive materials. The system employs the all-MiniLM-L6-v2 sentence-transformer for dense vector retrieval, whereas a cross-encoder re-ranking model refines candidate passages before answer generation. Two independent language models a locally hosted FLAN-T5 instance and the GROQ cloud API generate answers in parallel, with an automated selection mechanism choosing the more grounded response. The research consists of five major sections: introduction, analysis of relevant technologies, system requirements specification, system architecture description, system documentation, and testing. The thesis incorporates conclusions and references in its final sections. The scope of this work consists of 58 core pages, yet it contains 11 figures together with 6 tables, and 18 bibliographical sources as its foundation.