AWS Builder Center

AuditIQ Cloud: Building Evidence-Backed AI Financial Auditing with Amazon Bedrock & Strands Agents

Learn how AuditIQ Cloud combines Amazon Bedrock foundation models with deterministic Pandas execution via the Strands Agents SDK to deliver zero-hallucination, evidence-grounded financial audit investigations.

Introduction: The Challenge of AI in Financial Auditing

Corporate finance and internal audit teams process hundreds of thousands of transactions across multiple spreadsheets, ERP dumps, and vendor records. Traditional spreadsheet reviews are slow and manual. However, generic "ChatGPT for Excel" tools present severe risks: they hallucinate financial numbers, execute untrusted arbitrary code ( eval / exec ), and lack verifiable data lineage required for audit committee scrutiny.
To solve this trust deficit, our team built AuditIQ Cloud—an enterprise financial investigation platform deployed natively on AWS.
Our core design philosophy is simple:
"AI decides the investigation plan; deterministic code executes the math."



Architecture Overview

AuditIQ Cloud follows a modern serverless multi-tier architecture built on AWS.
┌──────────────────────┐
│ Auditor / User │
└──────────┬───────────┘
│ HTTPS
▼
┌──────────────────────┐
│ AWS Amplify Hosting │
└──────────┬───────────┘
│
├────────── Authenticate ──────────► ┌────────────────────┐
│ │ Amazon Cognito │
│ └─────────┬──────────┘
│ │
│ JWT Bearer Token
│ │
▼ │
┌───────────────────────────────────────────────────────────┴───┐
│ AWS App Runner (FastAPI Backend) │
│ │
│ Model Reasoning & Tool Selection │
└───────────────┬───────────────────────┬───────────────────────┘
│ │
│ │
▼ ▼
┌──────────────────────────┐ ┌──────────────────────────────┐
│ Amazon Bedrock │ │ Amazon S3 │
│ │ │ │
│ Claude 3 / Nova │ │ Raw Datasets & PDF Reports │
└──────────────────────────┘ │ Encrypted Storage │
└──────────────────────────────┘
┌──────────────────────────────┐
│ Amazon DynamoDB │
│ │
│ Metadata & Audit Runs │
└──────────────────────────────┘



How We Integrated Amazon Bedrock & Strands Agents SDK

Sending raw financial datasets into LLM prompts creates privacy leakage and context window bloat. Instead, AuditIQ Cloud uses Amazon Bedrock strictly for intent classification and tool selection using the Strands Agents SDK ( strands-agents ).

Step 1: Compact Schema Profiling

When a dataset (CSV/Excel) is uploaded, FastAPI generates a lightweight metadata profile containing column names, inferred data types, non-null counts, and sample distinct values—without sending raw records to the LLM.

Step 2: Intent Routing & Tool Selection

The user asks a natural language question such as:
"Are there any vendor duplicate invoices or split payments structured just below approval thresholds?"
The Strands Agent evaluates the schema profile and selects from pre-sanitized, allow-listed audit tools ( find_duplicate_vendors ,  detect_spending_spike ,  find_round_invoices , etc.).
# Sample Strands Agent initialization with Bedrock Model
from strands import Agent
from strands.models import BedrockModel
bedrock_model = BedrockModel(
model_id="anthropic.claude-3-haiku-20240307-v1:0",
region_name="us-east-1"
)
agent = Agent(
model=bedrock_model,
system_prompt="You are an expert financial audit agent. Select safe, allow-listed tools.",
tools=[find_duplicate_vendors, detect_spending_spike, find_split_payments]
)

Step 3: Safe Deterministic Execution

The tools run strictly using Pandas & NumPy against the actual records stored securely on serverless storage. No arbitrary code execution is permitted.



Key Results & Takeaways

  1. Zero Math Hallucinations: Because all calculations run in Pandas, numbers are 100% accurate.
  1. Data Privacy: Raw customer transactions are never used for LLM model training.
  1. Instant Scalability: Leveraging Amazon Bedrock via AWS App Runner allows the platform to handle complex natural language queries seamlessly.
By combining generative intelligence with deterministic execution, AuditIQ Cloud brings trust and speed back to financial auditing.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article