AWS Builder Center

Builder Center

Building IRONCLAD Sentinel: Autonomous Construction Billing Audit with AWS Strands Agents | Agents for Humans

Building IRONCLAD Sentinel: Autonomous Construction Billing Audit with AWS Strands Agents | Agents for Humans

How we built IRONCLAD Sentinel using AWS Strands Agents SDK and Amazon Bedrock AgentCore to autonomously audit commercial construction draw packets, eliminate LLM math hallucinations, and prevent statutory Prompt-Pay penalties.

Jaguri

What We Built

Commercial construction finance bleeds billions annually to retainage calculation discrepancies, statutory Prompt Payment Act interest penalties (1.5%–2% monthly), and mechanics lien fraud. Standard conversational chatbots are a catastrophic failure mode in this space—they hallucinate financial arithmetic and fail at chronological date reconciliation.
For the AWS Agents for Humans Hackathon, we engineered IRONCLAD Sentinel: an institutional-grade, zero-chat executive verification engine that executes forensic financial audits behind the scenes and halts at an authenticated 1-click Human-in-the-Loop decision gate.

Architecture & AWS Implementation

IRONCLAD is orchestrated as a multi-agent Directed Acyclic Graph (DAG) leveraging AWS native tools:
  1. AWS Strands Agents SDK: Built using a high-throughput Tri-Track fan-out/fan-in topology:
    • ForensicAuditSentinel (Professional Track): Extracts draw packet line items and executes deterministic retainage and lien waiver chain-of-custody checks.
    • FairPayStatutoryGuardian (Good Neighbor Track): Evaluates subcontract rider clauses against a 14-jurisdiction statutory prompt-pay reference table.
    • EverydayDecisionCardEmitter (Everyday Track): Synthesizes upstream data into an executive decision card without free-text chat noise.
  2. Amazon Bedrock AgentCore Deployment: Containerized with native agentcore.yaml and Dockerfile ready for deployment on Bedrock AgentCore runtime, utilizing Claude 3.5 Sonnet (us.anthropic.claude-3-5-sonnet-20241022-v2:0) for semantic clause classification and Claude 3.5 Haiku for execution.
  3. Zero-LLM Deterministic Math Invariant: Financial math is strictly isolated from LLM token generation. Retainage withholding and net disbursements are computed with 100% Python Decimal precision.
  4. Executive Interface & Observability: Real-time state transitions stream via non-blocking FastAPI Server-Sent Events (SSE) to a Next.js 16 dark-theme console. Telemetry is traced with OpenTelemetry GenAI semantic conventions exported to AWS CloudWatch and Langfuse.

Key Learnings

In enterprise compliance, the role of LLMs must be strictly bounded: generative inference should handle noisy, unstructured document perception, while financial math, statutory day counting, and payment authorization must remain 100% deterministic and enforced by code-level guardrails.
# python# architecture
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