ComplyAct: Governed Agentic AI for Auditable Automation

The Problem: The "Hallucination Abyss"

Enterprise automation is currently trapped between two failing extremes:

1. Legacy RPA: Brittle, expensive to maintain, and "blind" to human nuances like handwritten notes.
2. Black-Box AI Agents: Powerful but unpredictable. In compliance-heavy industries (Finance, Healthcare, Logistics), a 1% hallucination isn't a bug—it’s a multi-million dollar liability.

The result? 80% of enterprise AI pilots fail to reach production because "Trust cannot be audited."

The Solution: ComplyAct

ComplyAct provides a third way: Governed Agentic Automation. We’ve turned "Rogue AI" into a trusted digital workforce by engineering a system that knows exactly when to ask for help.

Key Innovations:

• Multimodal Ingestion (Powered by Nova Pro): Seamlessly parses messy, scanned, and handwritten documents that stop traditional OCR in its tracks.
• Graceful Halt Logic (Governance-First): When the agent encounters low-confidence data (<80% score), it doesn't guess. It halts and escalates to a human expert via a Slack-integrated Decision Modal.
• Hyper-Sync Engine (Nova Act Simulation): A zero-latency WebSocket layer that synchronizes AI "thoughts" with "Ghost Cursor" UI actions at 2x speed, narrating every step in a Transparency Terminal.
• Cryptographic Finality (Amazon QLDB): Every AI decision and human override is hashed into an immutable ledger, transforming software actions into permanent legal records.

How We Built It (The AWS Stack)

We utilized the full power of the Amazon Nova ecosystem to create a seamless, end-to-end orchestration:

• Brain: Amazon Nova Pro for high-reasoning multimodal document extraction.
• Hands: Amazon Nova Act reasoning patterns implemented via a deterministic Playwright executor for robust UI navigation.
• Voice: Amazon Nova 2 Sonic for real-time, low-latency voice approvals of "Graceful Halts," allowing executives to govern workflows hands-free.
• Ledger: Amazon QLDB (Quantum Ledger Database) to store the immutable audit trail of the human-AI collaboration.
• Orchestrator: A high-performance FastAPI backend coupled with a Next.js 14 "Iron-Man HUD" frontend.

Prize Category Qualifications

• ✅ Best Agentic AI Application: ComplyAct is a true agent. It ingests data, forms a plan, recognizes its own confidence limits, and interacts with external UIs.
• ✅ Best UI Automation Application: We use Nova Act patterns to dynamically map and navigate legacy ERP interfaces, turning unstructured PDFs into structured UI actions.
• ✅ Best Use of Amazon Nova: We integrated the entire stack—Nova Pro (Vision), Nova Act (Action), and Nova 2 Sonic (Voice)—into one unified workflow.
• ✅ Enterprise Governance (IBM Alignment): Every interaction is immutably logged, solving the primary security and compliance concerns of Fortune 500 companies.

Challenges We Overcame

• The Latency War: Synchronizing a high-reasoning LLM with a live UI automation without "lag" was our biggest hurdle. We solved this by building a WebSocket-driven Hyper-Sync layer that buffers AI thoughts and plays them back in a choreographed sequence.
• Multimodal Ambiguity: Designing the "Graceful Halt" threshold required deep testing of Nova Pro's confidence scoring to ensure the AI stops for a coffee stain but ignores a simple smudge.

Accomplishments We're Proud Of

• Successfully processing a document with handwritten marginalia that 3 other top-tier LLMs failed to extract correctly.
• Building a UI that makes "Complex Backend Reasoning" transparent and beautiful to a non-technical auditor.

What's Next for ComplyAct

• Multi-Agent Swarms: Deploying specialized agents for different ERP modules (HR, Finance, Legal) that coordinate through a central "Governance Master."
• Predictive Compliance: Training models on the "Human Overrides" to proactively flag potential compliance risks before they enter the system.

Built With

• act
• amazon
• fastapi
• next.js
• nova
• playwright
• pro
• qldb
• sonic
• supabase
• tailwindcss
• websockets