Date of Submission

2026-10-23

Supervisor

Engr. Dr. Syed Irfan Nabi, Associate Professor, Department of Computer Science

Committee Member 1

Dr. Asad Arfeen, Examiner – I, NED University, Karachi

Committee Member 2

Dr. Faisal Iradat, Examiner – II, Institute of Business Administration (IBA), Karachi

Degree

Master of Science in Computer Science

Department

Department of Computer Science

School

School of Mathematics and Computer Science (SMCS)

Keywords

Fintech, Regulatory Technology, Retrieval-Augmented Generation, Large Language Models, Governance Risk and Compliance

Abstract

The rapid expansion of fintech through digitalized financial services is associated with both enhanced speed and scale, but also the development of increasingly complex architectures. Fintech organizations have limited capability in governance, risk, and compliance (GRC) due to the predominantly manual and retrospective nature of their current GRC processes. In response to these challenges, this research adopts a Design Science Research methodology to design, develop, and preliminarily assess the feasibility of an Enterprise Compliance Framework (ECF), a web-based, multi-step agent pipeline for regulatory gap analysis. The ECF classifies the intent of a user query, performs section-aware document chunking, uses standard-specific Retrieval Augmented Generation (RAG) to ground compliance statements in the applicable regulatory standard, employs structured planning and execution, provides authenticated per-user document isolation, maintains persistent agent memory, enforces citation constraints, and logs all activities. Currently, the ECF regulatory knowledge base contains 18 regulatory frameworks, including Pakistan-specific sources such as State Bank of Pakistan (SBP) regulations and the Prevention of Electronic Crimes Act (PECA) 2016. An expanded benchmark of 42 compliance questions was used, covering single-fact, multi-standard, and multi-step categories, with seven questions held back from tuning and used as a robustness check. Seven configurations were compared, each given a similar amount of retrieved text. Two judge models scored every answer without being told which configuration produced it. The use of retrieval showed an increase in average scores of 1.03 points from a total possible score of 5 (95% confidence interval [0.62, 1.44]) compared with not using it. Additionally, retrieval showed better grounding of citations: while all 71 regulatory identifiers given by the model were without support when there was no retrieval, only 9.4% of the identifiers provided were without support under the best configuration. When retrieval was used, it produced the correct source document for every question, at a recall@5 of 1.00. The best configuration used standard-specific retrieval with fixed-size chunking and produced a score of 4.17, notably higher than both a single retrieval query at 3.97 and no retrieval at 2.94. However, the evaluation did not show that either section-aware chunking or the multi-step planner improved accuracy. The study reports the system as a functional prototype. Before treating these findings as settled, further expert assessment and testing in a live compliance environment will be required.

Document Type

Restricted Access

Submission Type

Thesis

Download

Share

COinS