Skip to content

Platform

A fraud intelligence layer, built in parts that can be reasoned about.

The platform is being designed as a set of components with clear responsibilities: deterministic controls the institution owns, models that add judgement, analytics that expose structure, and a workspace where a person reaches a decision that can be defended afterwards. What each component produces is described here; how they are combined internally is not.

Detection
Rules, ML, anomaly, graph
Output
Score, evidence, recommendation
Decision
Authorised bank personnel
Record
Versioned and auditable

The fraud intelligence engine

Four disciplines evaluate the same transaction, each contributing a different kind of reading, and each keeping the reasoning that produced it so the final position can be explained rather than asserted.

Four independent evaluations of the same transaction, resolving into one position.

What the institution receives
  • Risk score
  • Contributing evidence
  • Plain-language explanation
  • Recommended action

How these are sequenced, combined and weighted is our own work, and is not published. What we do publish is the output — one risk position, the evidence behind it, and a person who decides.

Capabilities

What each part of the platform is responsible for, and what it is designed to produce.

01

Rules engine

Deterministic controls the institution owns, evaluated ahead of any model.

  • Bank-defined thresholds, lists and conditions
  • Versioned rule sets with change history
  • Rule outcomes carried into the final explanation
  • Designed so a rule can be added without a model release
02

Machine learning

Supervised models that score fraud probability from historical outcomes.

  • Contribution to the score retained for every event
  • Training on institution-specific labelled outcomes
  • Candidate models validated before any production release
03

Anomaly detection

Deviation measured against an account's own established behaviour.

  • A behavioural baseline per account, rather than one shared threshold
  • Detection that does not depend on a labelled example existing first
  • Designed to reduce dependence on known fraud examples
04

Graph intelligence

Accounts, devices, beneficiaries and paths analysed as a network.

  • Entity resolution across identifiers
  • Community and cluster detection
  • Relationship density and transaction pathways
05

Risk decisioning

One consolidated risk position with a recommended course of action.

  • One consolidated position across every detection method
  • Configurable thresholds for alerting and prioritisation
  • Recommendation, not automated customer action
  • Alert prioritisation designed around analyst capacity
06

Explainability

The evidence behind a score, in reviewable language.

  • Signal-level contribution to the final score
  • Plain-language indicators for each contributing factor
  • Evidence retained with the case record
  • Written for investigators, reviewers and auditors alike
07

Case management

The workflow from alert to documented outcome.

  • Queues, assignment and status tracking
  • Investigation notes and attached evidence
  • Outcome capture with reason codes
  • Designed to fit existing fraud team process
08

Model governance

Control over what is deployed, and when.

  • Versioning across models, features and rules
  • Performance and drift monitoring
  • Validation gates ahead of release
  • Institutional approval in the release path
09

Audit & reporting

A reconstructable record of how each decision was reached.

  • Immutable action and decision trail
  • Model version attached to every scored event
  • Operational and case reporting
  • Built for internal audit and supervisory review

The workspace

Interface concepts for the teams who would work the queue: dense, traceable, and organised around the decision rather than around the chart.

TasawurAI/Risk overview
Open alerts

34

+6 last hour

In review

12

4 analysts

Escalated

5

2 to compliance

Closed today

21

9 confirmed

Alert volume by hourLast 24h
00:0012:0023:00
By typology
  • Account takeover38%
  • Mule network24%
  • Velocity19%
  • Structuring12%
  • Other7%
ReferenceTimeAmountChannelScoreBandStatus
TX-8271914:22:0712,400.00Mobile82HighOpen
TX-8270414:19:513,150.00Web74HighIn review
TX-8268814:16:33890.00Mobile61ElevatedIn review
TX-8267114:11:0222,000.00Branch58ElevatedOpen
TX-8265014:04:48410.00Mobile44ElevatedAssigned
TX-8261213:58:191,275.00Web27LowClosed

Interface concepts in development. All references, amounts and scores shown are sample data.

Built for accountable AI.

Every capability described here is being designed against the same constraint: a regulated institution has to be able to explain, monitor and reconstruct any decision the platform contributed to.