Use cases · Anomaly detection · Financial crime
Explain the block — to the customer
and the regulator.
A customer’s transaction history is a behavioural sequence. Senua AI learns each account’s normal states and flags the transaction that genuinely departs — with the state, the departure and the confidence attached. Every automated decision arrives with its reasoning, inside your own environment.
The scenario
3,000 reviews a week — and the score that failed model-risk review.
A mid-size bank’s rules engine queues three thousand reviews a week. Analysts clear them at forty seconds each and still miss the layering case that makes the papers. The ML vendor’s score cut the queue but failed the model-risk review: nobody could explain a single decision, the model needs quarterly retraining, and the data must not leave the country.
Explainability of automated decisions is now a compliance requirement, not a nice-to-have — and a black-box score cannot meet it, no matter how accurate.
How Senua AI handles it, end to end
Behavioural states per account, learned continuously, decided with an audit trail — in your datacentre.
1 · Learn each account’s normal
The engine induces behavioural states from the account’s own history — amounts, velocity, counterparty patterns — alongside the book’s population states. No feature pipeline to maintain.
2 · Decide with a decision file
A flagged transaction shows the learned state, the departure, and the confidence trajectory — a decision file, not a score. That is what survives a model-risk and regulator audit.
3 · Absorb new behaviour in-session
A customer’s legitimate new salary pattern is absorbed as it appears — the false positive stops without a retraining cycle or a model release.
Why it’s different
Fewer reviews, every one explained, data residency by construction.
The queue shrinks to signals.
Confidence-graded departures replace binary rule trips — analysts investigate behaviour that genuinely changed, not thresholds that happened to trip.
Model risk becomes tractable.
The model documentation is the reasoning path each decision carries. There is no separate explainability layer to bolt on and defend.
Nothing leaves your environment.
Runs in the bank’s own datacentre on commodity CPU — no third-party model API, no data egress, no GPU line item.
An honest caveat.
Public fraud datasets are partly synthetic and heavily imbalanced. We frame public-data results against each dataset’s published baselines, and claims about a real book are made only after an audit of that book — never from public data alone.
The public proof — run it yourself
Validated on public labelled data, in the open.
We validate this use case on public labelled transaction corpora — PaySim (mobile-money with labelled fraud) and the ULB credit-card fraud set — with a pre-registered protocol: learn behavioural states from clean history, replay the labelled windows blind, measure detection against the labels and the false-positive rate against published baselines.
Start with your own book.
A retrospective audit on your historical transactions, inside your environment: we show you which known cases surface as behavioural departures — and at what review-queue cost.