Implementing Proof of Concept of Private Generative AI for Audit Automation

TECH STACK:
AWS Bedrock|Azure AI Studio|Private GPT|Python
SERVICES:
AI/ML

Dedicated team behind the project

PARTNERSHIP

The Client

PROJECT SCOPE

The Challenge

STRATEGIES AND EXECUTION

What Was Done

To ensure we delivered the best solution, we evaluated three potential setups before recommending Azure OpenAI Studio. Each option was analyzed based on its ability to meet the project’s key requirements of security, scalability, and efficiency:

/ 01: Ubuntu Server with PrivateGPT

This setup allowed maximum control and customization, particularly in selecting models and embeddings for data processing. However, it required significant maintenance and was complex to manage, making it more suitable for larger projects with dedicated engineering teams. It was not the right fit for InvestiNet’s immediate needs due to high maintenance costs.

/ 02: AWS Bedrock

AWS Bedrock provided a stable and highly reliable environment with minimal bugs and an easy setup process. It allowed us to experiment with vector storage for optimized data retrieval. However, the platform offered a limited selection of models compared to other solutions, which could be a drawback for projects requiring a wide range of options.

/ 03: Azure OpenAI Studio

Azure OpenAI Studio stood out for its maturity and extensive model availability, including high-performance GPT models from OpenAI and other industry leaders like Microsoft and Meta. The platform’s integration capabilities and user-friendly interface made it a strong candidate, especially for U.S.-based projects where data security and compliance are critical.

Implemented Features:

feature:/ 01
LLM Integration

Integrated Azure OpenAI Studio’s Large Language Models (LLMs) to automate responses for security audit questionnaires, ensuring high accuracy and reducing manual effort.

feature:/ 02
Custom API Integration

Developed custom API solutions to seamlessly connect the audit automation tool with Investinet’s existing systems, ensuring smooth data flow and minimal disruption to current workflows.

feature:/ 03
Data Encryption and Secure Data Handling

Implemented strong data encryption protocols to protect sensitive audit information, in line with U.S. regulatory standards, including HIPAA and data privacy laws.

feature:/ 04
Model Fine-Tuning and Optimization

Fine-tuned GPT models within Azure to align with Investinet’s specific audit requirements, optimizing for both speed and accuracy.

feature:/ 05
Implementation of Local LLM on Different Infrastructures

We implemented a local version of the LLM to give InvestiNet full control over their data processing environments. This deployment was configured across multiple infrastructures, including on-premises servers and cloud platforms.

ACHIEVEMENTS

Results

Potential for 50% Time Reduction in Audit Processes

The PoC showed that by automating the completion of security audit questionnaires, Investinet could reduce the time senior management spends on audits by up to 50%, potentially saving around $18,000 annually

30-40% Potential Reduction in Operational Costs

The PoC indicated that switching to a serverless architecture could reduce operational costs by 30-40%, primarily by eliminating the need for extensive infrastructure maintenance and dedicated engineering teams

Full Compliance with U.S. Data Security Regulations

The PoC verified that Azure’s platform is fully capable of meeting U.S. data security and compliance standards, including HIPAA, ensuring secure handling of sensitive audit data

Scalable Solution for Future Growth

The PoC confirmed that Azure OpenAI Studio’s serverless design can support Investinet’s future growth without significant changes to infrastructure, ensuring scalability for evolving business needs

TECHNOLOGIES IN USE

Tech Stack

Daniel

Head of Engineering at Ralabs

Simplify Your Audits with AI

Let our experts help you save time, reduce costs, and ensure compliance with a customized AI solution

Relevant Cases

You got it right!

Only 21% of people can identify an accessible visual.

your question