• 

  homescreen

Inspiration

Every month, people sit down with their bank statements and scan line by line, double checking charges and searching for anything unfamiliar, hoping nothing slipped through. It is time consuming, stressful, and reactive. We were inspired to build a system where trust is automatic and fraud is stopped before users ever have to worry about it.

What it does

Cryos is not a traditional fraud detection system. It does not rely on global averages or static rules that treat every user the same. Instead, each agent learns the behavior of an individual account over time, building a personalized understanding of what is normal for that specific user. When suspicious activity is detected, Cryos does not immediately block access or react blindly. The system reasons through possible causes, analyzes detailed account context, and deploys honeypot traps to safely observe malicious behavior. Agents collaborate to trace the source of the threat and anticipate potential next actions. Based on this reasoning, Cryos plans appropriate responses, logs meaningful evidence, and determines the safest next steps. When necessary, the system escalates by notifying the account holder and guiding preventive actions to stop fraud before damage occurs.

How we built it

Cryos uses a multi-agent architecture with 20 specialized AI agents orchestrated by a MasterOrchestrator that coordinates analysis through a WorkflowEngine managing learning and attack phases. Each agent extends BaseAgent and uses OpenRouter API calls with tiered model selection (GPT-4o-mini for detection, Claude Haiku for specialized tasks, Claude 3.5 Sonnet for strategy), making structured JSON responses through callLLM() and callLLMWithTools() with function calling support. The system maintains persistent state through TransactionMemory (tracking transaction history, balance, device history, and baseline patterns), MemoryStore (episodic, semantic, procedural, and working memory), and VectorMemory (RAG retrieval via semantic search over document embeddings), with MemoryRepository persisting account state across sessions. Agents access context through callLLMWithMemory(), which bundles recent transactions, baseline patterns, episodic memory, procedural rules, and RAG-retrieved documents into context prompts, while ToolCollection enables agents to execute tools (log_note, update_learning_memory, query_transaction_history, etc.) with rate limiting and validation, creating a multi-step reasoning loop where agents can iteratively call tools, retrieve relevant historical patterns, and synthesize findings into fraud detection decisions that evolve through meta-learning from each transaction analysis cycle.

Challenges we ran into

One of the biggest challenges was balancing accuracy with speed. Preventing fraud requires fast decisions without generating false positives that disrupt legitimate transactions. Designing coordination between multiple agents while keeping the system efficient was also a significant challenge. While we worked on efficiency for a good bit, the agents still take a significant time to run the complete workflow.

Accomplishments that we're proud of

We successfully built a working multi agent fraud prevention system that operates continuously and removes the need for manual transaction reviews. We are especially proud of creating a solution that prioritizes user peace of mind while maintaining strong security guarantees. Additionally, building a agent (Kratos) to commit fraud for testing was a new experience for us.

What we learned

We learned that proactive security is far more effective than reactive detection. Building with automation and intelligent agents from the start leads to systems that are more scalable, reliable, and user friendly.

What's next for Cryos

Next, we aim to partner closely with banks to create a better, more seamless experience for every customer they serve. Cryos is close to being production ready and designed to operate at scale. By minimizing false alerts and eliminating unnecessary friction, it delivers security that works quietly in the background. Our goal is to help banks move beyond reactive fraud prevention and toward a future where customers feel protected without ever being interrupted.

Built With

• agenticai
• claude
• css
• express.js
• framer
• html
• multiagentaisystem
• node.js
• openai
• openroutersdk
• react
• three.js
• vapi
• vite
• websocket