# Superview

> Superview (by Serverless AI, Inc.) is an intent firewall for AI agents in banks and fintechs. It checks every agent action against the whole conversation before it runs, in under 0.8 seconds, and returns allow, step-up or pause.

## What it stops
- Outward: scams that manipulate customers through an AI agent — authorized push payment (APP) fraud, fake bank or police staff, investment scams, money-mule recruitment, social engineering.
- Inward: the bank's own agent acting beyond the user's intent or its role permissions — out-of-scope lookups, third-party data in replies, excessive call volumes.

## Key facts
- Under 0.8 s per decision (general LLMs: 1.3–9.5 s on a 20K-token input; Artificial Analysis, Nov 2026).
- Stops 89% of FraudBench attacks (internal testing); best general model tested: 64.5% (Gemini 3.6 Flash).
- Works with any LLM; deploys in the customer's own cloud; data stays in your environment.
- Audit trail per decision; supports human oversight and intervention before execution, as expected by bank AI guidance such as the HKMA's.

## Research
- [FraudBench paper](https://arxiv.org/abs/2608.18136): 150 bank fraud scenarios for AI agents, built on Sierra's τ-Knowledge bank-agent benchmark.

## Getting started
- Free 7-day risk scan: 30-min call, connect in shadow mode (nothing blocked), then a report on what Superview would have stopped.

## Team
- Lu Xian, co-founder (product and research), ex-Alibaba Cloud
- Dheeraj Pai, co-founder, Carnegie Mellon

## Where to start without a customer-facing agent
- Staff AI: copilots and coding agents
- Scam checks on service chat and call transcripts
- Customers' own AI agents logging in to the bank

## Contact
- [Book a 30-minute call](https://cal.com/dheerajpai)