SWIVEL Guardian

Inspiration

Many fraud systems are designed to answer one question: "Did someone unauthorized make this transaction?"

But some of the most damaging scams happen when the real account owner willingly sends the money because they are being manipulated. A scammer may impersonate a bank, government agency, family member, or trusted organization and pressure the victim with urgency, threats, or secrecy.

We built SWIVEL Guardian around a different question:

"Are you making this transaction because someone is manipulating you?"

What it does

SWIVEL Guardian is an AI-powered scam intervention layer designed to sit inside a bank or credit-union payment flow.

When a customer attempts a transaction, Guardian first examines behavioral signals such as:

transaction amount
recipient history
whether the recipient is new
device and geographic context
normal transfer behavior

If the payment looks unusual, Guardian does not immediately deny it. Instead, the payment enters a temporary review state and an AI agent begins an adaptive conversation with the customer.

The agent asks contextual questions to understand why the payment is being made. It looks for social-engineering indicators such as urgency, secrecy, threats, external instructions, or impersonation.

Guardian keeps two kinds of evidence separate:

Behavioral evidence:
"This is a new recipient and the amount is far outside the customer's normal range."

Conversation evidence:
"The customer said a government representative threatened arrest and told them not to contact their bank."

If the situation remains suspicious, Guardian escalates the case to a bank employee with an explainable summary. The AI can recommend review or escalation, but the human employee retains final authority over high-risk financial decisions.

How we built it

The application includes a customer-facing banking prototype, an AI intervention system, and an employee fraud-review dashboard.

Our architecture separates two types of intelligence:

A behavioral risk engine determines whether a transaction is abnormal for that customer.
A Gemini-powered intervention agent investigates the customer's intent and identifies potential social-engineering patterns.

The agent uses structured outputs rather than a simple chatbot response so that detected scam signals, recommendations, confidence, and explanations can be displayed clearly to both the customer and the bank employee.

We also designed the behavioral risk layer behind a provider interface so a statistical anomaly-detection model can be integrated without changing the rest of Guardian.

Trusted recipients are supported as well, but trust does not automatically bypass fraud protection. Even a trusted recipient can trigger intervention if the transaction is dramatically outside the customer's normal behavior.

Challenges we ran into

One of our biggest design challenges was avoiding an unrealistic assumption that the bank could simply read a customer's private texts, emails, or phone calls.

Instead, Guardian uses only information available within the financial environment and discovers the social-engineering context by talking directly with the customer.

Another challenge was balancing safety with customer autonomy. We did not want an AI model to permanently block access to someone's money. Guardian therefore uses an intentionally asymmetric authority model: the AI can pause a suspicious payment, release a pause when an explanation is clearly legitimate, or recommend human review, but cancellation remains a human decision.

We also had to design for false positives. An unusual transaction is not automatically fraudulent, which is why behavioral anomaly detection triggers an investigation rather than an accusation.

Accomplishments that we're proud of

We built a complete end-to-end flow where:

transaction anomaly → AI investigation → social-engineering detection → explainable case → human review

Our live Gemini agent generates adaptive follow-up questions instead of following only a fixed questionnaire, and behavioral evidence remains visibly separate from conversational evidence throughout the system.

We also built the system so that the behavioral-risk component can be replaced by an external statistical model through a clean API without rewriting the application. :chatgpt-content-reference{index="0"}

Our core scam scenario demonstrates something important: even when the correct customer is authenticated, their device is known, and their location looks normal, Guardian can still detect that an unusual payment may be the result of social engineering and escalate it for human review. :chatgpt-content-reference{index="1"}

What we learned

We learned that detecting unusual financial behavior and understanding human intent are two different problems.

Statistical models are good at recognizing when a transaction deviates from normal behavior. Large language models are better suited to understanding explanations, persuasion, threats, secrecy, and other social-engineering signals.

Combining those capabilities while keeping a human in the final decision loop created a much stronger system than relying on either approach alone.

What's next

The next step is to connect Guardian to a larger behavioral-risk model trained on richer transaction histories and evaluate false-positive and scam-detection performance at scale.

We also want confirmed employee outcomes to feed back into the system so future risk models and investigations can learn from previously verified cases.

Long term, Guardian could operate as an API-based scam-intervention layer integrated into payment platforms such as SWIVEL and the financial institutions they serve.