Anomalies caught as they happen
Kubli Pulse watches your recurring flows and flags what departs from normal behaviour, before the operation completes. It works on payments, insurance claims, utility usage and returns: the model learns from your own closed cases, not an industry average.
Each row is a movement arriving. Almost all of them flow through without anyone noticing. When the behaviour departs, the row lights up and the operation stops before it completes.
No hand-written rule says «block anything over €2,500 at night». The model learned what an ordinary day looks like at your institution and recognises what does not resemble one.
Confirmed fraud cases are rare, and they stay locked in the file because they contain names and account numbers. That is why almost every system on the market learned on somebody else's fraud.
Sectors
Not just payments
Insurance claims
A claim that departs from the customer's or policy type's normal pattern, flagged before payout.
Utility consumption
Usage that doesn't match that meter's usual readings, caught before it turns into undetected loss.
E-commerce abuse
A return, a promo code use, or an account that behaves differently from the store's usual traffic.
Vendor payments
An invoice that duplicates an order already fulfilled, or an amount outside that vendor's usual pattern.
How it works
Kubli Pulse watches your recurring flows and flags what departs from normal behaviour, before the operation completes. It works on payments, insurance claims, utility usage and returns: the model learns from your own closed cases, not an industry average.
Your confirmed anomaly files are the most useful material you own, and the one thing that never leaves your systems. Echo makes them usable with the customer taken out, so the model recognises your patterns instead of an industry average.
In nearly every other treatment the outlier is noise to be damped. Here it is the signal: flatten the anomalies and there is nothing left to recognise. Anonymisation protects who the customer is and leaves the behaviour untouched.
The block happens inside your perimeter, on the actual operation. On an anonymised file you could not stop any account, because you would not know which one: the treated material is for training, not for deciding.
What weighed on the decision, which version of the model made it, and who overrode it if anyone did. You need this the day you stop the wrong one and the customer asks why.
The other modules
Usable archives, measured risk
Kubli Echo produces a usable copy of your archives and measures its re-identification risk with the three attacks the EDPB names: singling out an individual, linking them to another dataset, inferring an attribute. Above the threshold, the file is not released.
Discover EchoDocumented automated decisions
Kubli Compass records what every automated decision about a person was based on, at the moment it is made. It produces the documentation the EU AI Act requires for high-risk systems, including differences in behaviour between groups.
Discover CompassUsable logs, anomalies surfaced
Kubli Sentinel anonymises access and system logs, so you can keep them, share them and analyse them. It then filters the stream and surfaces only what departs from a normal day on your network.
Discover SentinelTwo routes
If you have a problem to solve, start with a demo on your own case. If you are assessing us, as a fund or a procurement team, start with the confidential documentation.
Kubli issues no certifications and holds none. It hands over the measurements taken and the documents to stand behind them when someone asks.
Thirty minutes. You tell us which data is blocked and what you would do with it, and we show you the result on an example close to yours. If the case is not a fit we say so on the call.