About the Project – Sentinel
What inspired us
Financial cybercrime is evolving faster than traditional rule-based systems can keep up. Fraudsters use mixer chains, darknet wallets, and automated techniques to exploit blind spots in financial systems. We wanted to build something that doesn’t just react but continuously thinks, adapts, and improves on its own.
With powerful tools from TRM Labs, Skyflow, Redis, and Anthropic, this hackathon felt like the perfect opportunity to create an autonomous, production-grade fraud defense system that could actually help banks, crypto exchanges, and fintech apps stay ahead of attackers.
What we built
We created Sentinel, an autonomous, self-improving fraud guard powered by four cooperating agents:
🟦 Watcher Agent
Monitors incoming transactions, identifies high-risk events, and decides when deeper investigation is needed.
🟧 Detective Agent
Uses TRM Labs intelligence, Redis vector search, and learned playbooks to evaluate blockchain wallets, detect dark-web patterns, and produce a risk assessment with evidence.
🟥 Guardian Agent
Makes real enforcement decisions—block, step-up authentication, escalate, or allow—and triggers real-time actions to protect users and systems.
🟩 Coach Agent
Evaluates past decisions using ground truth, computes metrics, and creates new fraud playbooks to improve system performance over time.
🔒 Privacy & Trust
All customer PII is tokenized using Skyflow, ensuring privacy and regulatory compliance.
🧠 Memory & Intelligence
Every decision, embedding, case, cluster, and playbook is stored in Redis long-term memory, enabling similarity matching and continuous learning.
How we built it
Frontend (Lovable + Next.js) • Built a complete Sentinel dashboard with views for Live Investigation, Case Details, Playbooks, Agent Metrics, and Activity Log. • Designed a professional UI that visualizes each agent’s reasoning steps and actions.
Backend (Cursor + Node.js + TypeScript) • Implemented REST APIs for investigations, metrics, playbooks, and event logs. • Integrated Redis JSON + RediSearch for case storage and vector similarity. • Implemented autonomous multi-agent orchestration with Anthropic’s Claude models. • Created realistic TRM and Skyflow clients and later swapped them into the real API flow. • Built a self-improvement loop with APS scoring (Agent Performance Score).
What we learned • How to design multi-agent architectures with clear roles, handovers, and autonomy. • How to use Redis not just as a database, but as a long-term memory for agents, with vector search powering behavioral similarity. • How to leverage TRM Labs to detect illicit activity and dark-web exposure. • How to apply Skyflow tokenization to protect customer data without losing context. • How to make agents explainable, generating human-readable evidence traces. • How to close the loop between decisions and outcomes to enable self-improving AI systems.
Challenges we faced • Designing an agent system that is both autonomous and safe, ensuring Guardian Agent decisions are explainable. • Structuring Redis data models for: • cases • embeddings • event logs • metrics • playbooks • agent performance tracking • Getting vector search tuned to produce meaningful similarity results for dark-web patterns. • Prompt engineering Claude for structured reasoning and reliable JSON outputs. • Managing the handoff between 4 different agents while keeping logs, metrics, and evidence synced. • Balancing real-time responsiveness with intelligent decision-making in less than a second.
Why this matters
Cyber fraud is becoming autonomous. Defenses need to be too. Sentinel shows how financial institutions can deploy production-grade AI agents that: • detect emerging threats • defend automatically when needed • escalate intelligently • explain every action • and continuously get smarter
all without ever exposing sensitive customer data.
Log in or sign up for Devpost to join the conversation.