Software Engineering - KPMG (NDA)

Overview

I engineered a Retrieval-Augmented Generation (RAG) pipeline to automate audit Quality Assurance (QA), bridging the gap between cutting-edge AI and the rigorous compliance standards of the Financial Services industry.

Learning and Growth

Over the summer, I joined KPMG’s Financial Services practice as a software engineering intern. I was placed on a team centered around modernizing legacy audit workflows and reducing the latency of document verification for large-scale banking and insurance clients. The system I worked on enabled Audit Managers and Field Teams to instantly query and verify vast amounts of financial documentation, emphasizing precision and data integrity. While specific client portfolios and proprietary risk models are under NDA, I can share the engineering challenges I navigated and the solutions I architected to reach our goals.

Takeaway

Building a production-grade NLP tool for financial services involved moving beyond theoretical models to handling the "messiness" of real-world enterprise data. The first phase of my project focused on dissecting the manual QA process. This meant collaborating with Senior Auditors and Financial Analysts to understand where human error was most prevalent and defining the specific "high-risk" documents that needed automated scrutiny. Though I initially relied on my manager to navigate these cross-functional requirements, I eventually took ownership of the technical specifications for the data ingestion layer.

A large portion of my time was spent optimizing the RAG pipeline using Python and SQL. I didn't just want the model to work; I needed it to work within KPMG’s secure infrastructure. I experimented with different LLM frameworks and embedding strategies to handle complex financial terminology. I learned how to prioritize system reliability over pure speed, realizing that in a financial audit context, accuracy is non-negotiable. After multiple iterations of the retrieval logic, I built a scalable PyTest suite to rigorously validate the system, eventually achieving 95% precision in our automated benchmarks.

The experience was technically demanding, pushing me to understand the rigor required for enterprise software engineering in a regulated industry.

My internship with KPMG was a deep dive into the intersection of AI and financial services. I was solving a problem where "hallucinations" (AI errors) could have significant compliance consequences. Navigating the integration of modern Python microservices with legacy SQL and Enterprise backends was a steep learning curve. Overall, I learned that successful software engineering isn't just about writing complex code—it's about building robust, testable systems that solve a tangible business problem, and I’ve become much more confident in shipping code that drives operational efficiency.