SentinelFin

💡 Inspiration Over the last two years, global regulators handed down nearly $7 Billion in Anti-Money Laundering (AML) fines—including a historic $3.09 Billion penalty against a single bank for systemic compliance failures.

When I looked into why this happens, the answer was always the same: analyst burnout. To avoid fines, banks hire armies of compliance analysts who spend several hours manually jumping between disconnected legacy silos, rigid CSV exports, and cryptic SWIFT messages just to draft a single Suspicious Activity Report (SAR).

I realized that simply bolting a basic generative AI chatbot onto existing bank data wouldn't solve this crisis. I needed to build a system capable of true enterprise-grade reasoning, backed by strict regulatory governance.

⚙️ What it does SentinelFin completely reinvents the AML investigation pipeline, reducing the drafting time of a legally-complex SAR from several hours to under 5 minutes.

It starts when a user drops massive, unstructured financial documents into my custom dashboard. SentinelFin instantly parses the data and routes it to my massive 8-agent backend. These agents work in tandem: filtering entities against FBI watchlists, crawling offshore shell company networks, and dynamically writing custom Python code to detect money laundering patterns.

However, AI should never make final legal decisions autonomously. Therefore, SentinelFin is entirely governed by a UiPath Maestro Case. Maestro natively suspends the AI pipeline at critical decision points and generates native UiPath Action Center tasks. Only when a human Triage Analyst or BSA Officer logs in and clicks "Approve" does Maestro permit the AI to finalize the FinCEN Form 111 payload and sync the audit trail to UiPath Data Service.

🛠 How I built it To achieve this, I utilized a true "Boss and Worker" architecture:

The Boss (Governance): I used UiPath Maestro as the top-level orchestrator. It natively handles the high-level Case Plan lifecycle, SLA timers, and Action Center human-in-the-loop gating. The Worker (Reasoning): I built a Directed Acyclic Graph (DAG) of eight highly specialized AI agents using Python, FastAPI, and LangGraph, powered by Meta Llama 3.1 (70B) via AWS Bedrock. The Co-Pilot (Coding Agents): To bridge the gap between my external Python AI and the UiPath ecosystem, I utilized the UiPath Antigravity CLI as an autonomous coding agent. It acted as my developer co-pilot, scaffolding the complex LangGraph state-machine and writing the exact Python OData payloads required to interact with the UiPath Cloud REST APIs. The Presentation (Transparency): I built a React/Vite dashboard that utilizes Server-Sent Events (SSE) to stream the AI's real-time "thought process" and workflow status directly back to the user. 🚧 Challenges I ran into Integrating a fully autonomous, high-speed LangGraph pipeline with an asynchronous enterprise platform as a solo developer was incredibly challenging.

⚠️ The Staging Lab Pivot: In my video demonstration, you will see native UiPath Action Center handling the Human-in-the-Loop approvals flawlessly. When deploying to the final live Staging Lab, I discovered my specific tenant did not have the Action Center service provisioned. In any other enterprise platform, this would be a dead end. Instead, it highlighted my absolute favorite feature of UiPath Maestro: its extreme architectural flexibility! To ensure judges could still fully interact with my live deployment, I was able to instantly pivot and engineer a custom HTTP Polling architecture right inside Maestro. When you test my solution in the lab, Maestro natively pauses and waits for you to click "Approve" at my custom React terminal. I have never seen a platform that provides such powerful, diverse options to overcome roadblocks!

🏆 Accomplishments that I'm proud of I am incredibly proud of my live Server-Sent Events (SSE) streaming architecture. Watching the React dashboard update in real-time as the 8 agents pass memory and massive data payloads to each other provides a level of transparency that is critical for trust in the financial sector.

I am also extremely proud of how I made the finalized compliance data instantly accessible. By integrating UiPath Autopilot, compliance teams can interact with closed cases natively. Instead of digging through raw JSON payloads or 5-page legal narratives, a user can instantly ask Autopilot to summarize a finalized SAR or query specific historical transaction details, dramatically accelerating post-investigation audits.

📖 What I learned I learned the immense value of combining Agentic AI frameworks with Robotic Process Automation (RPA). AI is brilliant at unstructured reasoning and dynamic code generation, but it lacks the guardrails, auditability, and deterministic routing that enterprise software requires. By combining LangGraph with UiPath Maestro, I created a highly resilient, enterprise-ready solution.

🚀 What's next for SentinelFin I am expanding SentinelFin from a reactive investigation tool into a proactive, global defense grid.

First, I am designing a Federated Learning Consortium built on UiPath Data Service. This will allow multiple global banks to securely share anonymized transactional topologies and emerging laundering patterns without compromising customer privacy, creating a unified defense against financial crime.

Second, I am expanding my Dynamic Coding Agent to execute Autonomous Asset Freezing. In the future, the exact millisecond a BSA Officer signs the final SAR in Action Center, SentinelFin will use UiPath RPA to immediately execute API calls across decentralized ledgers and international banking systems to freeze illicit funds in real-time, stopping criminals before the money can cross international borders.