About the prototype

Know your agent
before it pays.

An explanatory note on what SENTINEL KYA does, how it detects fraud, and the technology behind it.

01

Why it exists

The line between people and bots is blurred. AI agents can browse, compare and pay on a customer's behalf, and many of them are legitimate. At the same time, fraudsters use scripts and hijacked agents that look exactly like those helpful assistants. A rule that blocks all automation stops genuine customers; a rule that trusts all automation opens the door to mule accounts and account takeover.

SENTINEL KYA (Know Your Agent) answers one question for every payment session: who is acting, are they allowed to do this, and what should the bank do next? It returns one of four actions: allow, step-up (ask for extra proof), hold for a person to review, or block, always with plain-language reasons.

02

What the prototype does

The prototype is a working proof of concept with eight screens. The Trust Desk lets a risk analyst describe an agent, its mandate and a transaction, and receive an AI-assisted assessment. The Capital Engine turns fraud prevented into Basel IV capital impact. The Fraud Lab runs ten synthetic agents, from a retail customer to a forged-signature agent, through the engine and shows step by step how each fraud is caught. The Live Engine scores raw session event logs, streams ten thousand sessions, and seals each decision into a tamper-evident audit trail. The Technology page shows the architecture, the Build Tracker shows progress against each build phase, and the Pitch Deck page holds the investor presentation.

03

How fraud is detected

Detection happens in four steps. First, the engine reads behaviour: the time between events, how regular that timing is, whether anyone typed, how many clicks or API calls occurred, and whether the browser reports automation. Humans are irregular; scripts are machine-regular; AI agents tend to move with even, deliberate pauses and no typing.

Second, it classifies the actor as a human, a scripted bot or an AI agent. Third, if an agent is acting, it checks the signed mandate the customer gave it: is the signature valid, is the mandate still in date, is the payee on the approved merchant list, is the amount within the cap, and has the daily limit been reached? Fourth, it combines these into a trust score from 0 to 100 and maps it to an action. Hard mandate rules always win: a failed rule can make a decision stricter but can never make it more lenient.

04

The two decision paths

The live path is deterministic. It uses fixed rules and behaviour features, with no AI model and no network call, so every decision is reproducible and completes in well under a millisecond — far inside the 100 ms budget a payment flow allows.

The analyst path, on the Trust Desk, calls a large language model through the Lovable AI Gateway with zero data retention. The model reads the agent profile and scenario and returns a structured verdict: agent class, four weighted trust pillars (delegation, scope, behaviour, provenance), reason codes and fraud typologies. The same hard mandate rules are then applied on top, so the model can explain but never overrule bank policy.

05

Technology stack

The application is built with React 19 and TanStack Start, a full-stack framework that renders pages on the server and runs back-end logic as typed server functions. It is styled with Tailwind CSS using a strict Swiss red-and-white design system, and deployed on an edge runtime close to users. Every decision is stored in a managed Postgres database (Lovable Cloud) protected by row-level security: anyone may submit a well-formed assessment, but a signed-in person can only read the records they created.

06

Modern Swiss banking architecture

The design follows how modern Swiss banks are moving away from monolithic core systems and overnight batch runs. Each capability is a decoupled service behind small RESTful APIs. Events flow through a streaming backbone modelled on Apache Kafka, with three topics: incoming session events, decisions, and audit records. In the prototype, ten thousand sessions stream through in milliseconds, and on a realistic synthetic traffic mix 99% are decided straight through with no human touch; only holds reach a person.

Client secrecy is protected by tokenisation: payee and session identifiers are replaced with HMAC-SHA-256 tokens before anything is recorded, so the original values are never stored. Each audit entry is sealed with a SHA-256 hash that includes the previous entry, forming a chain. Editing any past record breaks the chain, and the Live Engine demonstrates this with a Try tampering button.

The same decision contract is designed to sit in front of either major Swiss core platform: Avaloq, whose portfolio-object model suits multi-currency wealth management, or Temenos Transact, whose arrangement-based ledger is built for high-volume transaction processing.

07

Basel IV alignment

Basel IV, in force in Switzerland since January 2025 under FINMA Circular 2023/1, asks banks to measure operational risk, prove the lineage of the data behind decisions (BCBS 239), and keep critical services resilient. SENTINEL KYA is designed to support those duties: every decision is explainable, traceable to its inputs, sealed against tampering and fails safe to step-up rather than silently allowing a payment. It is designed for alignment, not a certification.

08

Swiss Basel IV and the Capital Engine

Switzerland applies Basel IV through the Capital Adequacy Ordinance and five FINMA ordinances: CreO (credit risk), MarO (market risk), LROO (operational risk), TBEO and DisO (reporting and disclosure). The prototype's Capital Engine is a first: it links fraud stopped at the payment session to the bank's operational-risk capital. Blocked fraud never enters the loss database, which lowers the Loss Component and the Internal Loss Multiplier under the Standardised Measurement Approach. The same screen runs internal and standardised credit RWA side by side, applies the output floor (up to 72.5%), and traces every number to its source in BCBS 239 style, down to the hash-chained KYA decision.

09

What is real today, and what comes next

Running today: behaviour detection, agent classification, mandate rules, AI-assisted analysis, the streaming backbone, tokenisation, the hash-chained audit trail, audit export and persistent storage. Pilot targets that need a bank's own infrastructure: hardware security modules holding the keys, AES-256 encryption at rest under bank control, mutual TLS between services, live Avaloq or Temenos adapters, and a sovereign Swiss cloud deployment. The next milestone is a twelve-week pilot on a bank sandbox with signed agent mandates, measured against the go/no-go scorecard in the pitch deck.

Team Quantum Leap · Raja Mukherjee · Priyanshu Dhar · Aditi Kulkarni