New Titles

Deep Learning with Python

Chollet, Francois

1- WHAT IS DEEP LEARNING? 2- THE MATHEMATICAL BUILDING BLOCKS OF NEURAL NETWORKS 3- INTRODUCTION TO TENSORFLOW, PYTORCH, JAX, AND KERAS 4- CLASSIFICATION AND REGRESSION 5- FUNDAMENTALS OF MACHINE LEARNING 6- THE UNIVERSAL WORKFLOW OF MACHINE LEARNING 7- A DEEP DIVE ON KERAS 8- IMAGE CLASSIFICATION 9- CONVNET ARCHITECTURE PATTERNS 10- INTERPRETING WHAT CONVNETS LEARN 11- IMAGE SEGMENTATION 12- OBJECT DETECTION 13- TIMESERIES FORECASTING 14- TEXT CLASSIFICATION 15- LANGUAGE MODELS AND THE TRANSFORMER 16- TEXT GENERATION 17- IMAGE GENERATION 18-BEST PRACTICES FOR THE REAL WORLD 19- THE FUTURE OF AI 20- CONCLUSIONS. Deep Learning with Python, Third Edition puts the power of deep learning in your hands. This new edition includes the latest Keras and TensorFlow features, generative AI models, and added coverage of PyTorch and JAX. Learn directly from the creator of Keras and step confidently into the world of deep learning with Python.In less than a decade, deep learning has changed the world—twice. First, Python-based libraries like Keras, TensorFlow, and PyTorch elevated neural networks from lab experiments to high-performance production systems deployed at scale. And now, through Large Language Models and other generative AI tools, deep learning is again transforming business and society. In this new edition, Keras creator François Chollet invites you into this amazing subject in the fluid, mentoring style of a true insider.

Date Published:2026

Writing Secure Code : Practical strategy and techniques for secure application coding in a network world

Howard, Michael

Keep black-hat hackers at bay with the tips and techniques in this entertaining, eye-opening book! Developers will learn how to padlock their applications throughout the entire development processfrom designing secure applications to writing robust code that can withstand repeated attacks to testing applications for security flaws. Easily digested chapters reveal proven principles, strategies, and coding techniques. The authorstwo battle-scarred veterans who have solved some of the industrys toughest security problemsprovide sample code in several languages. This edition includes updated information about threat modeling, designing a security process, international issues, file-system issues, adding privacy to applications, and performing security code reviews. It also includes enhanced coverage of buffer overruns, Microsoft® .NET security, and Microsoft ActiveX® development, plus practical checklists for developers, testers, and program managers.

Date Published:2023

Zaikamate : ai-powered interactive cooking assistant

Tariq, Habiba [22L-7501]

Submitted in fulfillment of the requirements for the degree of BS in Data Science to the Department of Data Science. List of Figures List of Tables 1 Introduction 1.1 Purpose of this Document 1.2 Intended Audience 1.3 Definitions, Acronyms, and Abbreviations 1.4 Conclusion 2 Project vision 2.1 Problem Domain Overview 2.2 Problem Statement 2.3 Problem Elaboration 2.4 Goals and Objectives 2.5 Project Scope 2.6 Sustainable Development Goal (SDG) 2.6.1 Industry, Innovation, and Infrastructure 2.6.2 Responsible Consumption and Production 2.7 Constraints 2.8 Business Opportunity 2.9 Stakeholders Description / User Characteristics 2.9.1 Stakeholders Summary 2.9.2 Key High-Level Goals and Problems of Stakeholders 2.10 Conclusion 3 Literature Review / RelatedWork 3.1 Definitions, Acronyms, and Abbreviations 3.2 Detailed Literature Review 3.2.1 SuperCook 3.2.2 Cookpad 3.2.3 Tasty 3.2.4 Yummly 3.2.5 Paprika Recipe Manager 3.2.6 SideChef 3.2.7 Whisk 3.2.8 Kitchen Stories 3.2.9 ChefTap 3.2.10 Allrecipes Dinner Spinner 3.2.11 Zesto 3.3 Literature Review Summary Table 3.4 Conclusion 4 Software Requirement Specifications 4.1 List of Features 4.2 Functional Requirements 4.2.1 Functional Requirements for Users 4.2.2 Functional Requirements for System 4.3 Quality Attributes 4.4 Non-Functional Requirements 4.4.1 Availability 4.4.2 Security 4.4.3 Usability 4.4.4 Performance 4.4.5 Serviceability 4.5 Assumptions 4.6 Use Cases 4.7 Hardware and Software Requirements 4.7.1 Hardware Requirements 4.7.2 Software Requirements 4.8 Graphical User Interface 4.8.1 Sign up Screen 4.8.2 Login Screen 4.8.3 Homepage Screen 4.8.4 Enter Ingredient Screen 4.8.5 View Recipes List Screen 4.8.6 Recipe Detail Screen 4.8.7 Recipes Steps Screen 4.8.8 Avatar Screen 4.8.9 Share Recipes/Tips Screen 4.9 Database Design 4.9.1 ER Diagram 4.9.2 Data Dictionary 4.10 Risk Analysis 4.10.1 User Perspective Risks 4.10.2 System Perspective Risks 4.11 Conclusion 5 High-Level and Low-Level Design 5.1 System Overview 5.1.1 Recipe Fetching Module 5.1.2 Ingredient-Based Search 5.1.3 AI Avatar Interaction 5.1.4 Real-time Data Communication 5.1.5 User Experience and Personalization 5.2 Design Considerations 5.2.1 Assumptions and Dependencies 5.2.2 General Constraints 5.2.3 Goals and Guidelines 5.2.4 Development Methods 5.3 System Architecture 5.3.1 React Native (User Interface) 5.3.2 Backend (FastAPI Server) 5.3.3 Database 5.3.4 LLM 5.3.5 RAG-Based Agent 5.3.6 Speech-to-Text Engine (Transcription) 5.3.7 Text-to-Speech 5.4 Architectural Strategies 5.4.1 Moving with Efficient Programming Frameworks 5.4.2 Error Detection and Recovery 5.4.3 Generalized Approaches to Control 5.4.4 Database Abstraction and Data Persistence 5.4.5 Concurrency and Synchronization 5.5 Domain Model/Class Diagram 5.6 Policies and Tactics 5.6.1 Coding Guidelines and Conventions 5.6.2 Testing and Quality Assurance 5.6.3 Version Control and Traceability 5.6.4 Data Persistence and Management 5.6.5 Directory and Code Organization 5.6.6 Maintainability and Future Updates 5.7 Conclusion 6 Implementation and Test Cases 6.1 Implementation 6.1.1 Mobile Frontend Implementation 6.1.2 Text-Based Recipe Generation 6.1.3 Voice-Enabled Avatar Interaction 6.1.4 Integration of Modules with FastAPI 6.1.5 Database Implementation 6.2 Conclusion 7 Conclusions 7.1 Summary of Progress 7.2 Challenges and Limitations 7.2.1 Challenges 7.2.2 Limitations 7.3 Recommendations for Future Work 7.4 Future Vision 7.5 Conclusion. With artificial intelligence (AI) changing day-to-day life in many areas, the food and cooking space is now going through a transformation as well. The food trend of cooking at home led to the demand for digital assistants that help make cooking and food easier, more personal recipe suggestions and a focused approach to sustainable food practices. ZaikaMate comes in the market as a new AI-based mobile cooking assistant, which will transform the manner in which people cook and interact with food. It is culturally authentic, hands-free and interactive support to the home cooks to make the kitchen more efficient and enjoyable. The motivation for ZaikaMate arises from the constraints of existing recipe apps, which frequently prioritise Western cuisines, lack cultural relevance, and rely primarily on static text-based instructions. ZaikaMate is an integration of Large Language Models (LLMs), Speech-to-Text (STT), Text- to-speech (TTS), and Retrieval-Augmented Generation (RAG) that can assist in a contextually adaptive and handsfree manner. This enables the users to concentrate on cooking rather than navigation of the screen which will enable a more flowing, dynamic and immersive experience. Our project scope includes designing and developing a mobile application that can be used as an AIbased cooking assistant, allowing users to enter ingredients and get customized Pakistani recipes and provide interactive and step-by-step voice-guided instructions via a conversational AI representative. Cooking is a collaborative as well as a culturally interactive process where the users are able to share their recipes and tips in a community based setting. The main stakeholders are the home cooks, professionals, students, food lover and system administrators. The project is based on the studies that indicate that AI avatars can be very helpful in improving the learning and interaction with the user, which is why ZaikaMate is a culturally, as well as smart, platform that is necessary. The project is aligned with several Sustainable Development Goals (SDGs) including (SDG 9: Industry, Innovation, and Infrastructure) and (SDG 12: Responsible Consumption and Production). ZaikaMate fuses technology with tradition to promote food innovation, reduce food waste, and encourage sustainable cooking practices, all while improving the home cooking experience, learning, user engagement and accessibility. With respect to software requirements of ZaikaMate, we have defined both functional and non-functional requirements, where special emphasis has been put on security, usability, scalability and responsiveness of processes. ZaikaMate has advanced, intuitive recipe suggestions, real-time voice enabled guidance by the AI chef avatar, and social sharing of recipes. All of these features are incorporated into the system architecture and allow users to access personalized recipes, have step-by-step cooking guidance and have a hands-free and culturally authentic cooking experience. Finally,risk analysis determines possible technical, business, content and quality, user adoption. We have developed ZaikaMate using an Agile methodology, allowing for iterative development, continuous testing, and refinement based on user feedback to ensure adaptability to evolving user needs. The system is built with React Native for the frontend and FastAPI for the backend, managing complex real-time interactions. The architecture incorporates both high-level and low-level design, integrating AI modules, user interfaces, and databases to ensure efficient data flow—from ingredient input to recipe generation and interactive guidance. Optimized for low latency and scalability, ZaikaMate delivers responsive, real-time cooking assistance with a focus on user satisfaction and system performance. In the future, ZaikaMate suggests focusing more on the refinement of its AI-based interaction and the development of its recipe recommendation intelligence to make it more user-adaptive and faithful to the style of cooking. The further developments will be aimed at the personalization, voice interaction optimization, and the responsiveness of the avatar.Finally, ZaikaMate is envisioning to become a full-fledged AI kitchen assistant that will allow its users to cook more intelligently, eat better and live healthier and more sustainable lives through technological culinary advancement.

Date Published:2025

Planora

Nadeem, Ahmad [22L-7994]

Submitted in fulfillment of the requirements for the degree of BS in Software Engineering to the Department of Software Engineering. List of Figures List of Tables 1 Introduction 1.1 Purpose of this Document 1.2 Intended Audience 1.3 Definitions, Acronyms, and Abbreviations 1.4 Conclusion 2 Project Vision 2.1 Problem Domain Overview 2.2 Problem Statement 2.3 Problem Elaboration 2.3.1 Inefficient Process 2.3.2 Non-Serious Vendors and Fake Listings 2.3.3 Lack of Transparency 2.4 Goals and Objectives 2.5 Project Scope 2.6 Sustainable Development Goal (SDG) 2.7 Constraints 2.8 Business Opportunity 2.9 Stakeholders Description/ User Characteristics 2.9.1 Stakeholders Summary 2.9.2 Key High-Level Goals and Problems of Stakeholders 2.10 Conclusion 3 Literature Review / RelatedWork 3.1 Definitions, Acronyms, and Abbreviations 3.2 Detailed Literature Review 3.2.1 EventPro Management System 3.2.2 PlanIt Pakistan 3.2.3 Cvent Event Platform 3.2.4 EventMate Lahore 3.2.5 Eventbrite 3.2.6 ShadiBanao.pk 3.2.7 Whova Event App 3.2.8 Eventify Pakistan 3.2.9 SplashThat 3.2.10 ShadiSaga 3.2.11 EventPlan Hub Karachi 3.2.12 BriteBiz 3.2.13 Eventza Pakistan 3.2.14 Hopin Virtual Events 3.2.15 BookMyEvent.pk 3.3 Literature Review Summary Table 3.4 Conclusion 4 Software Requirement Specifications 22 4.1 List of Features 4.2 Functional Requirements 4.2.1 User Management 4.2.2 Event Browsing and selection 4.2.3 Vendor Profiles 4.2.4 Booking Management 4.2.5 Booking History 4.2.6 Notifications 4.2.7 Admin Panel 4.2.8 Optional Features 4.3 Quality Attributes 4.4 Non-Functional Requirements 4.4.1 Reusability 4.4.2 Performance 4.4.3 Reliability 4.4.4 Security 4.4.5 Scalability 4.4.6 Maintainability 4.4.7 Availability 4.4.8 Portability 4.4.9 Usability 4.5 Assumptions 4.6 Use Cases 4.7 Hardware and Software Requirements 4.7.1 Hardware Requirements 4.7.2 Software Requirements 4.8 Graphical User Interface 4.9 Database Design 4.9.1 ER Diagram 4.9.2 Data Dictionary 4.10 Risk Analysis 4.10.1 Technical Risks 4.10.2 Business Risks 4.10.3 Performance 4.11 Conclusion 5 High-Level and Low-Level Design 5.1 System Overview 5.2 Design Considerations 5.2.1 Assumptions and Dependencies 5.2.2 General Constraints 5.2.3 Goals and Guidelines 5.2.4 Development Methods 5.3 System Architecture 5.3.1 Subsystem Architecture 5.4 Architectural Strategies 5.4.1 Programming Language and Frameworks 5.4.2 User Interface Paradigm 5.4.3 Concurrency and Synchronization 5.4.4 Error Detection and Recovery 5.5 Domain Model/Class Diagram 5.6 Policies and Tactics 5.6.1 Conventions and Coding Guidelines 5.6.2 Coding Environment 5.6.3 System Testing 5.6.4 Maintenance of Software 5.6.5 Extensibility 5.7 Conclusion 6 Implementation and Test Cases 6.1 Implementation 6.1.1 User Authentication using JWT 6.1.2 Database Schema Integration 6.1.3 Customer Dashboard Implementation 6.1.4 Vendor Dashboard Implementation 6.1.5 Admin Dashboard Implementation 6.2 Test case Design and description 6.3 Test Metrics6.4 Conclusion 7 User Manual 7.1 Getting Started 7.1.1 Sign Up 7.1.2 Sign In 7.1.3 Forget Password 7.2 Customer Guide 7.2.1 Home Dashboard 7.2.2 Browsing Vendors 7.2.3 Making a Booking Request 7.2.4 Booking Status Tracking 7.2.5 Viewing Booking History 7.2.6 Viewing Upcoming Bookings 7.2.7 Viewing Future Bookings 7.2.8 Managing Favorite Vendors 7.2.9 Profile Management 7.3 Vendor Guide 7.3.1 Vendor Dashboard 7.3.2 Adding Services 7.3.3 Managing Services 7.3.4 Booking Status 7.3.5 Managing Booking Requests 7.3.6 Viewing Upcoming Bookings 7.3.7 Managing Service Slots 7.3.8 Profile Management 7.4 Admin Guide 7.4.1 Admin Dashboard 7.4.2 Vendor Approval Management 7.4.3 Managing Platform Users 7.4.4 Logout 7.5 Workflow Summary 7.5.1 Booking Workflow 7.5.2 Vendor Approval Workflow 7.6 Conclusion 8 Future Work 8.1 Conclusion 8.2 Future Work. Planora is created to solve big problems encountered in conventional ways of planning events and current platforms by offering a contemporary, online solution for arranging events efficiently. In the past, organizing an event meant people had to go to many vendors in person, ask lots of questions repeatedly and handle communication manually which caused delays and confusion. Present systems usually don’t have integration that allows users ease in comparing vendors or tracking bookings or managing several services at once. Additionally, customers often have difficulties finding trustworthy vendors that fit their budget, location and event needs. At the same time, vendors face challenges in reaching potential clients and managing their availability properly. For solving these problems, Planora offers an internet-based system which brings together customers, suppliers and managers on one platform. Those who are customers can sign up, look through event categories like weddings, birthdays or business events and sort vendors by type of service provided, cost range, rating or area. Suppliers have the ability to set up detailed profiles with information about their services , upload pictures or videos, update availability schedules and answer requests for bookings. The admin panel maintains safety and quality through letting administrators confirm vendor accounts, control user actions and observe platform statistics like overall bookings and income. The platform also includes Google Maps for finding vendors based on location and uses PostgreSQL for safe and effective data handling. Planora makes communication digital and booking processes automatic, which cuts down manual work. It also boosts clarity and improves the whole planning experience for people using it. The platform is built to be scalable, easy to maintain and adaptable for future upgrades ensuring long-lasting usability and constant betterment. To sum up, Planora reshapes the process of arranging events by providing a well-structured. clear. and handy online solution that is advantageous for customers as well as vendors and administrators.

Date Published:2026

Management information systems : managing the digital firm.

Laudon, Kenneth C.,

A case-based approach to IS in business today Management Information Systems: Managing the Digital Firm provides a comprehensive introduction to Information Systems that draws connections between MIS and business performance. The authors present real-world case studies that explain how well-known companies use IT to solve problems and achieve their objectives. This real-world approach helps students develop sought-after skills, learn to lead IS-related management discussions, and use IT to meet bottom-line results. The 18th Edition has been updated to cover important contemporary topics including the latest on AI, sustainability and analytics.

Date Published:

Power electronics : devices, circuits, and applications

Rashid, Muhammad H. (Muhammad Harunur), 1945-

Chapter 1 Introduction PART I Power Diodes and Rectifiers Chapter 2 Power Diodes and Switched RLC Circuits Chapter 3 Diode Rectifiers PART II Power Transistors and DC–DC Converters Chapter 4 Power Transistors Chapter 5 DC–DC Converters PART III Inverters Chapter 6 DC–AC Converters Chapter 7 Resonant Pulse Inverters Chapter 8 Multilevel Inverters PART IV Thyristors and Thyristorized Converters Chapter 9 Thyristors Chapter 10 Controlled Rectifiers Chapter 11 AC Voltage Controllers PART V Power Electronics Applications and Protection Chapter 12 Flexible AC Transmission Systems Chapter 13 Power Supplies Chapter 14 DC Drives Chapter 15 AC Drives Chapter 16 Introduction to Renewable Energy Chapter 17 Protection of Devices and Circuits Appendix A Three-Phase Circuits Appendix B Magnetic Circuits Appendix C Switching Functions of Converters Appendix D DC Transient Analysis Appendix E Fourier Analysis Appendix F Reference Frame Transformation Bibliography Answers to Selected Problems Index. This text covers the basics of emerging areas in power electronics and a broad range of topics such as power switching devices, conversion methods, analysis and techniques, and applications. Its unique approach covers the characteristics of semiconductor devices first, then discusses the applications of these devices for power conversions. Four main applications are included: flexible ac transmissions (FACTs), static switches, power supplies, dc drives, and ac drives.

Date Published:2018

Computer science illuminated.

Dale, Nell.,

Designed for the introductory computing and computer science course, the student-friendly Computer Science Illuminated, Eighth Edition provides students with a solid foundation for further study, and offers non-majors a complete introduction to computing. Fully revised and updated, the eighth edition of this best-selling text retains the accessibility and in-depth coverage of previous editions, while incorporating all-new material on cutting-edge issues in computer science. Authored by the award-winning team Nell Dale and John Lewis, the text provides a unique and innovative layered approach, moving through the levels of computing from an organized, language-neutral perspective.

Date Published:2023

Computer architecture : a quantitative approach

Hennessy, John L.,

Computer Architecture: A Quantitative Approach, has been considered essential reading by instructors, students and practitioners of computer design for nearly 30 years. The seventh edition of this classic textbook from John Hennessy and David Patterson, winners of the 2017 ACM A.M. Turing Award recognizing contributions of lasting and major technical importance to the computing field, along with new author Christos Kozyrakis, is fully revised with the latest developments in processor and system architecture. True to its original mission of demystifying computer architecture, this edition continues the longstanding tradition of focusing on areas where the most exciting computing innovation is happening, while always keeping an emphasis on good engineering design. Winner of a 2019 Textbook Excellence Award (Texty) from the Textbook and Academic Authors Association Each chapter follows a consistent framework: explanation of the ideas in each chapter; a "crosscutting issues" section, which presents how the concepts covered in one chapter connect with those given in other chapters; a "putting it all together" section that links these concepts by discussing how they are applied in real machine; and detailed examples of misunderstandings and architectural traps commonly encountered by developers and architects Includes "Putting It All Together" sections near the end of every chapter, providing real-world technology examples that demonstrate the principles covered in each chapter Covers new developments in GPU and CPU architectures, as well as domain specific architectures Features more comprehensive coverage of systems on chip and heterogeneity.

Date Published:2026

Medquick

Rashid, Zohaib [22L-7885]

Submitted in fulfillment of the requirements for the degree of BS in Software Engineering to the Department of Software Engineering. List of Tables 1 Introduction 1.1 Purpose of this Document 1.1.1 Key Objectives 1.2 Intended Audience 1.3 Definitions, Acronyms, and Abbreviations 1.4 Conclusion 2 Project Vision 2.1 Problem Domain Overview 2.2 Problem Statement 2.3 Problem Elaboration 2.3.1 Lack of Access to Healthcare 2.3.2 Long Waiting Time and High Costs 2.3.3 Inexistent integration in Existing Solutions 2.3.4 Trust and Safety Concerns 2.3.5 Affordability & Efficiency 2.3.6 Doctor's Perspective 2.3.7 Patient Expectations 2.4 Goals and Objectives 2.5 Project Scope 2.5.1 The system includes 2.5.2 The system excludes 2.6 Sustainable Development Goal (SDG) 2.6.1 SDG 3: Good Health and Well-Being 2.6.2 SDG 9: Industry, Innovation, and Infrastructure 5.2.1 Assumptions and Dependencies 5.2.2 General Constraints. 5.2.3 Goals and Guidelines 5.2.4 Development Methods 5.3 System Architecture. 5.3.1 Subsystem Architecture 5.4 Architectural Strategies 5.4.1 Technology Stack 5.4.2 Scalability and Adaptability 5.5 Class Diagram. Con 5.6 Policies and Tactics 5.6.1 Tools 5.6.2 Coding guidelines 5.6.3 User Interface. 5.6.4 Extensibility 5.6.5 System Testing 5.6.6 Maintenance 5.6.7 Algorithms 5.7 Conclusion 6 Implementation and Test Cases 6.1 Implementation 6.1.1 Backend Architecture, Firestore Integration, and Security 6.1.2 Flutter Frontend Implementation 6.2 Test Case Design and Description 6.3 Test Metrics 7 User Manual 7.1 Role Selection. 7.2 Sign-up (Patient only) 7.3 Login (All Roles) 7.4 Profile Completion 7.5 Consultation and Prescription Generation (Patient) 7.6 Order tracking and Notifications (Patient) 7.7 Review prescriptions (Doctor) 7.11 [12:42 PM, 10/7/2026] aasia majeed185@gmail.com: NATION 7.8 Create Delivery and Assign Rider (Pharmacy) 7.9 Accept Delivery and Perform Drop-off (Rider): 7.10 Monitor Platform and Manage Users (Administrator) 7.11 Logout (All Roles) 8 Conclusion and Work completed 8.1 Conclusion 8.2 Work Completed 8.2.1 Authentication and Authorization System 8.2.2 Backend and System Integration 8.2.3 AI Model Refinement and Training 8.2.4 Cloud Data Structuring and Storage 8.2.5 Medical Record and Document Handling 8.2.6 Pharmacy and Location-Based Enhancements.

Date Published:2026

Adlang

Ahmad, Muhammad [22L-7930]

Submitted in fulfillment of the requirements for the degree of BS in Software Engineering to the Department of Software Engineering. 1 Introduction 1.1 Purpose of this Document 1.2 Intended Audience 1.3 Definitions, Acronyms, and Abbreviations 1.4 Conclusion 2 Project Vision 2.1 Problem Domain Overview 2.2 Problem Statement 2.3 Problem Elaboration 2.4 Goals and Objectives 2.5 Project Scope 2.6 Sustainable Development Goal (SDG) 2.6.1 SDG 9: Industry, Innovation and Infrastructure 2.6.2 SDG 12: Responsible Consumption and Production 2.7 Constraints 2.8 Business Opportunity 2.8.1 Business Opportunity for Vendors 2.8.2 Business Opportunity for Platforms 2.9 Stakeholders Description/ User Characteristics 2.9.1 Stakeholders Summary 2.9.2 Key High-Level Goals and Problems of Stakeholders 2.10 Conclusion 3 Literature Review / RelatedWork 3.1 Definitions, Acronyms, and Abbreviations 3.2 Detailed Literature Review 3.2.1 Ad Auctions for LLMs via Retrieval Augmented Generation 3.2.2 Personalized Risks and Regulatory Strategies of LLM in Digital Advertising 3.2.3 Improving Generative Ad Text on Facebook using Reinforcement Learning 3.2.4 Chat-REC 3.2.5 Multi-Agent Conversational Recommender System (MACRS) 3.2.6 Beyond Retrieval: Generating Narratives in Conversational Recommender Systems 3.2.7 Ads in Conversations 3.2.8 Online Advertisements with LLMs: Opportunities and Challenges 3.2.9 Contextualizing Recommendation Explanations with LLMs: A User Study 3.2.10 AI-Driven Personalization in eCommerce Advertising 3.2.11 GenAI Advertising: Risks of Personalizing Ads with LLMs 3.2.12 Factors Influencing Artificial Intelligence Conversational Agents Usage in the E-commerce Field: A Systematic Literature Review 3.2.13 Towards a Middleware for Large Language Models 3.2.14 Truthful Aggregation of LLMs with an Application to Online Advertising 3.2.15 RARE Framework 3.3 Literature Review Summary Table 3.4 Conclusion 4 Software Requirement Specifications 33 4.1 List of Features 4.1.1 User Features 4.1.2 Vendor Features 4.1.3 Admin Features 4.2 Functional Requirements 4.3 Quality Attributes 4.3.1 Usability 4.3.2 Performance 4.3.3 Reliability 4.3.4 Security 4.3.5 Maintainability 4.3.6 Availability 4.4 Non-Functional Requirements 4.4.1 Performance 4.4.2 Reliability 4.4.3 Usability 4.4.4 Security 4.4.5 Maintainability 4.4.6 Availability 4.4.7 Scalability 4.4.8 Compatibility 4.5 Assumptions 4.6 Use Cases 4.7 Hardware and Software Requirements 4.7.1 Hardware Requirements 4.7.2 Software Requirements 4.8 Graphical User Interface 4.9 Database Design 4.9.1 Data Dictionary 4.10 Risk Analysis 4.10.1 Technical Risks 4.10.2 Ethical and Privacy Risks 4.10.3 Business Risks 4.11 Conclusion 5 Proposed Approach and Methodology 5.1 System Functioning Overview 5.1.1 Presentation Layer 5.1.2 Business Logic Layer (The Middleware) 5.1.3 Persistence Layer 5.2 Core Subsystems and Implementation 5.2.1 Vendor and Ad Management Subsystem 5.2.2 Ad Intelligence and Delivery Subsystem 5.2.3 User Interaction and E-commerce Subsystem 5.3 Development Methodology 5.4 Conclusion 6 High-Level and Low-Level Design 6.1 System Overview 6.1.1 Vendor Management Portal 6.1.2 Ad Intelligence Retrieval System 6.1.3 Incentive Scheme and Loyalty of User. 6.1.4 Real Time Insight and Analytics 6.1.5 Ethics And Privacy Protection 6.1.6 System Integration Layer 6.1.7 Admin Dashboard 6.2 Design Considerations 6.2.1 Assumptions and Dependencies 6.2.2 General Constraints 6.2.3 Goals and Guidelines 6.2.4 Development Methods 6.3 System Architecture 6.3.1 Rationale 6.3.2 User Interface 6.3.3 Backend 6.3.4 Database 6.3.5 Real-Time Features 6.3.6 AI Components 6.3.7 Subsystem Architecture 6.4 Architectural Strategies 6.4.1 Use of Node.js Middleware with LLM APIs 6.4.2 MongoDB for Flexible and Scalable Data Management 6.4.3 User-Centric Design and Ethical Ad Integration 6.4.4 Security and Privacy Considerations 6.4.5 Scalability and Modularity 6.4.6 Technology Stack Selection 6.5 Domain Model/Class Diagram 6.6 Architectural Policies and Tactics 6.6.1 Technology Stack Selection 6.6.2 API Communication Protocol 6.6.3 Version Control and Strategy of Branching 6.6.4 Software Testing Plan 6.6.5 The Style and Conventions of Coding 6.7 Conclusion 7 Implementation and Test Cases 7.1 Implementation 7.1.1 Presentation Layer 7.1.2 Business Logic Layer 7.1.3 Persistence Layer 7.2 Conclusion 8 Experimental Results and Discussion 8.1 Part 1: Identifying the Best System Prompt 8.1.1 Candidate System Prompts 8.1.2 Assumptions Made 8.1.3 Research Methodology 8.1.4 Evaluator LLM System Prompt 8.1.5 Sample Record 8.1.6 Results 8.1.7 Verification of Results 8.2 Part 2: Identifying the Best LLM 8.3 Conclusion 9 Conclusions 9.1 Conclusion 9.2 Future Work (Plan for FYP-2) 9.2.1 Creation of the Contextual Intelligence 9.2.2 Live LLM Connection 9.2.3 Finalizing the Analytics Dashboard 9.2.4 User Rewards System 9.2.5 Final Polishing and Testing: . Large Language Models (LLMs) are widely adopted in conversational systems, yet their monetization remains limited to subscriptions and unrelated advertisements. AdLang introduces a middleware framework that embeds contextually relevant ads into LLM responses naturally and transparently. Vendors submit policy-compliant ads stored in a central database, while user prompts trigger appropriate ad insertions. Users will also enjoy loyalty points that may be redeemed at the system shop. The sellers will be shown elaborate analytics regarding the campaigns. It is a combination of all these that forms a moral and scalable ecosystem of integrating the delivery of the ad, the user interaction, and the e-commerce into a single route that will be used to be able to create a sustainable stream of revenue necessary to keep the conversational AI platforms operating.

Date Published:2025

Multiple input diagnostic aid system (midas)

Saqib, Muhammad Huzyefah [22L-7916]

Submitted in fulfillment of the requirements for the degree of BS in Software Engineering to the Department of Software Engineering. Table of Contents List of Figures x List of Tables xii 1 Introduction 1 1.1 Purpose of this Document 1.2 Intended Audience 1.3 Definitions, Acronyms, and Abbreviations 1.4 Conclusion 2 Project Vision 2.1 Problem Domain Overview 2.2 Problem Statement 2.3 Problem Elaboration 2.4 Goals and Objectives . 2.5 Project Scope 2.6 Sustainable Development Goal (SDG) 2.7 Constraints 2.8 Business Opportunity 2.9 Stakeholders Description/ User Characteristics 2.9.1 Stakeholders Summary 2.9.2 Key High-Level Goals and Problems of Stakeholders 2.10 Conclusion 3 Literature Review / RelatedWork 3.1 Definitions, Acronyms, and Abbreviations 3.2 Detailed Literature Review 3.2.1 Symptomate: self-service symptom checker 3.2.2 CureMD 3.2.3 Buoy Health: check symptoms and find right Care 3.2.4 The Application of Artificial Intelligence in Medical Diagnostics: A New Frontier 3.2.5 A Continued Pretrained LLM Approach for Automatic Medical Note Generation 3.2.6 Medical errors and patient safety: Strategies for reducing errors using artificial intelligence 3.2.7 Application of large language models in disease diagnosis and treatment 3.2.8 Integrating Social Care into Healthcare: A Review on Applying the Social Determinants of Health in Clinical Settings 3.2.9 Developing a scalable FHIR-based clinical data normalization pipeline for standardizing and integrating unstructured and structured electronic health record data 3.2.10 Google DeepMind: MedGemma Technical Report 3.3 Literature Review Summary Table 3.4 Conclusion 4 Software Requirement Specifications 4.1 List of Features 4.2 Functional Requirements 4.2.1 User Management 4.2.2 Patient Repository Management 4.2.3 Medical Diagnosis Management 4.3 Quality Attributes 4.3.1 Performance 4.3.2 Scalability 4.3.3 Reliability 4.3.4 Accuracy 4.3.5 Interoperability 4.3.6 Security 4.3.7 Modularity and Extensibility Non-Functional Requirements 4.4.1 Performance 4.4.2 Usability 4.4.3 Scalability 4.4.4 Security 4.4.5 Compliance 4.5 Assumptions 4.6 Use Cases 4.7 Graphical User Interface 4.8 Database Design 4.8.1 ER Diagram 4.8.2 Data Dictionary 4.9 Risk Analysis 4.9.1 Technical and Development Risks 4.9.2 Business Risks 4.9.3 Data Privacy and Liability 4.10 Conclusion 5 High-Level and Low-Level Design 5.1 System Overview 5.1.1 Interactive User Interface 5.1.2 Clinical Asset Management 5.1.3 Asset Input Processing 5.1.4 Diagnostic Agent 5.1.5 Report Generator 5.2 Design Considerations 5.2.1 Assumptions and Dependencies 5.2.2 General Constraints 5.2.3 Goals and Guidelines 5.2.4 Development Methods 5.2.3 Goals and Guidelines 5.2.4 Development Methods 5.3 System Architecture 5.3.1 High Level Partitioning 5.3.2 Component Collaboration 5.3.3 High-Level System Architecture Diagram 5.3.4 Rationale for Decomposition 5.3.5 Subsystem Architecture 5.4 Architectural Strategies 5.4.1 Choice of Technologies 5.4.2 Architectural Strategies 5.4.3 Scalability and Adaptability 5.4.4 Data Management and Persistence Strategy 5.4.5 Concurrency and Synchronization 5.4.6 Error Detection and Recovery 5.4.7 Future Extension Strategy 5.5 Domain Model/Class Diagram 5.6 Policies and Tactics 5.6.1 Coding Guidelines and Conventions 5.6.2 Testing Strategy 5.6.3 Continuous Integration, Deployment and Version Control 5.6.4 Error Handling and Debugging 5.6.5 User Interface Design 5.6.6 Database and Data Management 5.7 Conclusion 6 Implementation and Test Cases 6.1 Implementation 6.1.1 Authentication 6.1.2 Patient Record Management 6.1.3 Diagnostic Engine 6.1.4 Clinical Transcription and Note Taking 6.1.5 Medical Imagery Processing 6.1.6 Lab Results Ingestion with OCR 6.1.7 Patient Report Generation 6.2 Test case Design and description 6.3 Test Metrics 6.4 Conclusion 7 User Manual 7.1 User Roles and Permissions 7.2 General User Operations 7.2.1 Login 7.2.2 Logout 7.2.3 Basic Navigation Tips 7.3 Clinical User Manual . The profession of medicine is challenged to a serious issue; a huge volume of patient information that is fragmented and distributed in silo systems such as electronic health records (EHRs). Imaging reports, lab results, and patient charts are often correlated manually by clinicians; this correlating process often makes them think much more. This discontinuity is one of the primary causes of diagnostic errors, which is a widespread and avoidable source of patient harm. It is approximated that misdiagnosis results in approximately 800,000 severe patient harms every year in the U.S., 371,000 of which are deaths and 424,000 irreversible disabilities. We suggest a solution to the irreplaceable gap in patient safety in the Multiple Input Diagnostic Aid System (MIDAS), a multi-agentic decision support system aimed at doctors. The MIDAS vision includes the development of the AI-based system that can combine, analyze, and present various patient data in a structured and actionable format. MIDAS will help clinicians, by integrating multimodal patient data, such as clinical notes, lab results, imaging, vital signs, and social determinants of health (SDOH), to develop explainable, workflow-integrated insights, which will decrease diagnostic delays and cognitive overload. The main part of the project is the creation of a multi-agentic system that is able to process these heterogeneous data streams. The individual agents will use a particular AI method, including natural language processing (NLP) of clinical notes, computer vision of medical images, and data analytics of lab data and vitals, to produce comprehensive patient data. The main system characteristics will be the production of prioritized differential diagnoses to facilitate prompt clinical decision-making process, provision of actionable recommendations based on the patient case, and the existence of a secure and longitudinal patient repository to support further care. The complete system will become available via a web application that will produce detailed and practical patient reports. MIDAS is not meant to replace clinicians, but as an effective aid to enhance their decision making process. Having automated data synthesis and offered insight prioritization, the platform will enable healthcare professionals to concentrate on critical thinking and patient care. The ultimate is to improve patient safety by ensuring that the risk of misdiagnosis that may have its roots in cognitive overload and human error will be reduced dramatically. Finally, MIDAS creates a more efficient, precise and secure diagnostic setting, which directly responds to the pressing demand of sophisticated, combined decision support systems in the contemporary healthcare.

Date Published:2025

Crash risk prediction in cryptocurrency markets using financial market and technical indicators : the moderating role of network maturity.

Hameed, Sarah Abdul,

This dissertation aimed to design an overall framework to predict crash risk in cryptocurency markets by integrating financial market variables, technical trading variables, and the moderating effect of the maturity of cryptocurrency networks. The research used a pane data approach to predict the effect of global financial market conditions and cryptocurrency market technical indicators like momentum patterns, volatility clustering, and downside asymmetry on downside risk in cryptocurrency market. Cryptocurrency network maturity index was constructed using on- chain activity metrics such as transaction volume and network efficiency. This study employec machine learning regression and classification algorithms to predict crash risk in the cryptocurrency market. The results highlighted the predictive strengths of different algorithms and enabled identification of variables with the highest predictive power. The findings are relevan to stakeholders by explaining why consideration of financial market fundamentals, as well as intrinsic cryptocurency market indicators is important in gauging crash risk. This dissertatior contributes to literature by conceptualizing cryptocurrency network maturity as a moderator ir crash-risk prediction models, and a multidimensional framework integrating traditional finance market microstructure, and blockchain analytics. Further directions include focus on deep-learning based prediction models for more sophisticated results, and exploration of different quantification of network maturity to further consolidate one of the main findings of this research, which is tha crash risk affects all cryptocurrencies synonymously, regardless of maturity levels.

Date Published:2026

Deep Iearning - based predictive models for forex market trends : practical implementation and profit performance evaluation.

Afaq, Muhammad,

Forex market is the biggest and the most volatile market of the world with a valuation of around 7.5 trillion USD. It is the most liquid financial market where currencies are traded twenty-four hours, five days a week. It is a decentralized market which operates through a network of banks, corporations and other institutions that trade directly or through brokers Researchers have long sought to develop methods and tools to predict the direction of a currency in the forex market. Initially these researchers developed statistical models to predict the direction of a currency in the forex market, which were replaced by AI models, nowadays. Forex history and macroeconomic data have been used to predict forex currency trends but no research so far incorporates such diverse data for forex prediction. Prior studies have been conducted on individual data groups in forex trend forecasting, this research study combines multiple sources across traditional, machine learning and deep learning models. We aim to combine multiple data sources to perform forex trend pre- diction. This study undertakes empirical evaluation of traditional, machine learning and deep learning models on multiple data sources and determines which models offer better economic value in market-based simulation constraints. This research is based on the time period from January 2018 to December 2024 for the currencies of USD, EUR, GBP, JPY, AUD, NZD, CHF and CAD from reputable global sources, including the IMF, World Bank, and ESG databases. Sentiment data has been obtained from the Global Knowledge Graph and oil volatility from the Energy Information Administration We find that the traditional models have higher accuracy (~70%) than machine learning and deep learning models whereas the Multi-Layer Perceptron (MLP) achieves higher economic performance (Sharpe ratio of 0.74 at ~62% accuracy), highlighting the need to align research motivation with practical implementation.

Date Published:2026

Predicting liquidity risk In SMEs through machine learning to improve financial stability.

Shahid, Usama Bin,

Liquidity risk is the most imminent financial challenge for small and medium-sized enterprises SMEs) and the most manageable risk affecting firm-level financial stability. This study designs and tests an early warning system for SME liquidity risk using an interpretable machine- learing model that overcomes key drawbacks of traditional liquidity risk assessment. Traditional assessment relies on a handful of static ratios evaluated only on a single balance sheet date and assumes that the relationships among these ratios are linear and backward-looking. The empirical benchmark for the analysis is the Polish Companies Bankruptcy Dataset, the public version of which contains 43,405 firm- year observations from manufacturing firms across five forecastins periods in a transition economy (2002 to 2013). Bankruptcy is used as a proxy for liquidity failure among manufacturing firms. The feature space is intentionally limited to 17 firm-level liquidity, related ratios, including solvency ratios, cash-flow coverage ratios, days-of-liquidity-cover ratios and working-capital ratios, plus five macroeconomic conditioning ratios. Four supervised classifiers are trained and evaluated independently for one- and three-year time horizons under stratified five-fold cross-validation; SHAP attribution is calculated at the one- and three-year time horizons. The ROC-AUC is highest for gradient-boosted trees, at 0.867 and 0.771, respectively for one- and three-year-to-failure, outperforming linear baselines by 5 to 10 percentage points Debt-coverage time, quick ratio, cash-flow-to-debt coverage, eamed-equity buffer, and days-of liquidity-cover are the most dominant signals, while macroeconomic indicators do not play a significant role. The results reveal that managers, lenders, and policymakers have a clear, forward looking early-warning tool for SME liquidity risk: liquidity ratios, with a lead time of two to thred years. The benefits and potential of a liquidity-restricted feature design within this framework together with its horizon-conditional interpretability, are described, and its transferability to SME: in other emerging markets, such as South Asia, is discussed.

Date Published:2026

A I- powered evaluation of crisis communication in aviation using its financial impact.

Qamar, Maryam,

This research fills the gap in the absence of a quantitative framework that is structurec and quantitative in assessing the quality of crisis communication in the aviation industry. Existing Studies have mainly concentrated on stakeholder responses or qualitative evaluations of communication during measurement of the content of official crisis statements. This study develops an AI- a Crisis, leaving a gap in the systematic powered framework that quantifies three main elements of communication, including tone, transparency and timing, and integrates them into a composite Post-Crisis Communication Score (PCCS) to address the gaps present. These components are quantified using transformer-based natural language processing models and semantic analysis on official airline statements issued during the first three days of a CrisiS. The framework is empirically tested with an event study approach methodology, where cumulative abnormal returns are computed to determine the magnitude of investor reaction. The regression analysis is then used to test the relationship betweer communication quality and market response. The results suggest that timing and transparency have a strong effect on the containment of the investor reaction, whereas the tone does not have a direct significant impact. The composite PCCS has higher explanatory power and can therefore be used as an evaluative benchmark. The paper concludes that the quality of structured communication can be quantitatively evaluated and connected to financial outcomes, both of which provide a theoretical advancement and a practical crisis management tool in the aviation industry.

Date Published:2026

Real estate price prediction : a machine learning approach using online property market data.

Khan, Muhammad Yaseen,

The purpose of this study is to develop a predictive model for real estate prices in Pakistan by integrating property characteristics that include structural features locational, and amenities in the property through using machine learning techniques The findings of this study aim to assist customers. investors. and policymakers ïnvestment and policy decisions by providing more accurate and transparent price predictions. in making informed property purchase, Real estate is one of the most important sectors contributing to the economy through its forward and backward linkages with around forty other industries. Therefore, building a machine learning based model that predicts house prices will also be very useful for the allied industries to plan their business activities and take their investment decisions. Moreover, this will also be useful for public departments such as federal board of revenue to make their revenue projections and policies. However, at present, the relevant data is mostly fragmented and investment activity is speculative. To cover this gap this study focuses to utilize a comprehensive dataset consisting of listings from zameen.com and predict house prices based on different features of houses like structural (number of rooms, bathrooms and amenities). geographical (location, city etc.) and property types houses flats etc.). For this, we utilized OLS model, Hedonic Model and different models of machine learning (ML) including Random Forest, XG Boost, Cat Boost and Gradient Boosting. ML models are applied nonlinearities and heterogeneity can be captured. In the methodological framework so that preprocessing, feature engineering, hyper parameter tuning and cross validation are performed so that the model is robust and performs wel. In findings the performance of machine leaming models showed better predictive performance in this study in terms of accuracy, predictive performance precision in which ensemble methods performs better t0 capture nonlinearities between different variables compared to the traditional econometric models. This study help: in providing key insights to consumer and stakeholders that are significantly important in decision making, pricing and investment planning it will also help in the formulation of policy for reducing speculation and improving the transparency in the real estate market. The future direction would involve by expanding the dataset to cover rental market and including images of property using deep learning models like LSTM for long term predictions and expanding sentiment model to social media by taking the indication from it in order to know about rea time market.

Date Published:2026

Leveraging machine learning models and explainable AI to enhance customer segmentation and marketing insights

Zafar, Musferah,

The rapid growth of e-commerce has led to an increase in customer data; however, traditional segmentation methods relying mainly on purchase behavior fail to capture deeper customer intent This study proposes a hybrid customer segmentation approach that integrates both browsıng behavior and purchase activity to generate more meaningful insights. A quantıtative and exploratory design is used in which several clustering algorithms appled and Integrated in an ensemble framework, such as K are Means, DBSCAN, Agglomerative, and Spectral Clustering o enhance the stability of segmentation. The dataset 1S analyzed on the OpenDataBay E- commerce dataset; the most important variables are the Ad CTR. Ad CPC. Ad Spend, Revenue, and Unıts Sold. SHAP is used to explain the model to increase the interpretability and aid in business decision-making The findings reveal a lear difterentiation of customer groups with different levels of engagement and purchasing behaviors, which is useful in targeting them, personalizing, and marketing effectively. The significance of this study lies in the fact that it offers a practical and interpretable framework of customer segmentation in e-commerce.

Date Published:2026

Predicting non- performing loans using machine and deep learning approaches : comparative evidence from Pakistan Indonesia and Malaysia

Asghar, Shayan,

This study aims to improve the prediction of non-performing loans (NPLs) in dua banking_systems where traditional statistical tools fail to capture the nonlinear high-dimensional relationships in modern banking data. It comparatively evaluates four models- Random Forest (RF), Artificial Neural Networks (ANN) Long Short-Term Memory (LSTM) and CNN-LSTM hybrid for Islamic and conventional banks in Pakistan, Indonesia, and Malaysia. Using bank financial data alongside macroeconomic variables, model performance was assessed or accuracy, precision, recall, and F1-score. Results show that Random Forest outperformed ANN for Islamic banks, achieving 96% accuracy, while ANN performed better for conventional banks (83-85%), reflecting the modeling challenges posed by risk-sharing structures and smaller samples in Islamic banking. Macroeconomic factors such as inflation and interest rates significantly influenced predictions, confirming their relevance to credit risk modeling. The findings imply that generic credit risk models applied uniformly across banking types are inadequate, and that institution- and region-specific predictive frameworks are required. The study adds value by delivering an interpretable framework-feature importance analysis identifies NPL ratios, capital adequacy, and Shariah compliance as key risk drivers- -that is directly actionable for credit managers, policymakers, and regulators, with future work directed toward synthetic data augmentation, hybrid RF-LSTM architectures. and SHAP-based explainability.

Date Published:2026

A neuro-symbolic fraud prevention framework for e-commerce transactions using spending profiles and rule-based purchase verification

Hira, Adina,

The dramatic growth in e-commerce has greatly enhanced market efficiency, market participation and consumer experience. But with this rise considerable financial risks have come to global online platforms, especially through ever-evolving fraud schemes. With the growth of digital economies, the methods of fraudsters are constantly evolving, making traditional fraud dctection approaches increasingly ineffective. Existing methods of fraud detection are broadly classified into rule-based and machine learning-based methods. Although rule-based approaches are straightforward and easy to understand, they are not adaptable to dynamic fraudulent patterns. In contrast, machine learning techniques enhance prediction accuracy but are often less explainable, less scalable, and struggle to effectively model complex relations in transaction data. Such shortcomings often lead to high false alarms and a lack of transparency in machine decision-making. This impacts botk efficiency and user confidence and poses regulatory challenges in the financial sector. This paper proposes a neuro-symbolic approach to fraud prevention systems, combining neural pattern recognition with symbolic reasoning, to ensure both performance and explainability. Our approach also integrates graph neural networks (GNNs) to capture relationships between users, devices accounts, and transactions. This allows fraudulent transactions to be detected effectively, including complex fraud schemes such as collusion and mule account networks, which are not often detected using traditional methods. Furthermore, behavioral spending patterns are correlated with symboli verification rules to collectively evaluate the authenticity of transactions, with fraud detectior decisions being both data-based and explainable. The system is evaluated on three public datasets showing its effectiveness. The proposed system achieves a fraud detection accuracy of96.8%, witt an improvement of 32% in false alarms over conventional machine learning models. Additionally it enhances the recall of fraud networks by 27%, demonstrating its effectiveness in detecting complex fraud networks. It is also robust to skewed transaction data distributions, which is typica. in real-world financial transactions. All in all, the proposed system offers a scalable anc explainable approach to fraud detection and can be further integrated with real-time biometric behavior monitoring systems and online learning for improved performance in large-scale onling transaction systems.

Date Published:2026

Machine learning-based credit scoring for microfinance institutions in Pakistan : enhancing loan decisions and financial inclusion

Ali, Muhammad Ghulam,

in developing countries, microfinance institutions (MFIs) face challenges in credit risk assessmen lue to lack of information on borrowers, quality of information, and scarce resources. In Pakistan his is further complicated by a credit risk assessment and management process that is largely manual and subjective, resulting in inconsistent loan approvals, increased default risk and altimately financial exclusion of the marginalised population. Even though machine learning (ML has been shown to significantly decrease credit risks in the global setting, empirical results on the end-to-end operationalization of machine learning in the Pakistani MFI market have been limited The thesis fills this gap by developing and analysing an extensive ML-based loan repayment prediction model in collaboration with Agahe Pakistan, using information. The paper presents a stringent data refinement flow that both standardizes actual historical loan-leve] heterogeneous MIS data and manages missingness, fixes inconsistencies, pehavioural, time, and contextual variables for local lending behaviour. It has a multi-tie and develops modelling approach, in which logistic regression is benchmarked against ensemble models, such as Random Forests, XGBoost, and LightGBM, and, where possible, shallow neural networks Based on institutional risk thresholds, model performance is evaluated using discrimination (AUC: ROC, PR-AUC), calibration (Brier Score, Expected Calibration Error), as well as operational lecision metrics (precision, recall, F1-score, false negative rate). Robustness checks determine the effects of class imbalance treatment, strategy of imputation, period validation, and stability of features across branches and loan products, as well as borrower segments. In order to promote responsible adoption of AI, such tools as fairness diagnostics and interpretability, such as SHAP- based explanations, comparison of subgroup performance, and monotonicity validation, are also included in the study. The results demonstrate that the ensemble models are superior to the linear baselines in terms of discriminatory ability and risk ranking as well, and they maintain a reasonable palibration under the condition of being assisted by well-organized pre-processing. Featurc stability analysis clearly shows that the demographic-only scoring is not appropriate approach; i reveals the predictive significance of behavioural and repayment-pattern indicator. The fairness analysis uncovers subgroup disparities and enables proactive risk management, in line witt responsible lending guidelines. This thesis provides Pakistan-specific empirical evidence, links calibration with fairness - a gap in previous studies - and presents a scalable ML credit scoring model for resource-limited microfinance institutions. Practically, it offers MFIs a practical guide o enhance portfolio quality, mitigate default risks and promote financial inclusion with a transparent, auditable, and contextual model.

Date Published:2026