8 AI Engines Working in Concert

Each engine analyses a different dimension of the NDIS ecosystem. Together, they create an integrity intelligence layer that no single-engine system can match.

01 Network Graph 02 Behavioural Drift 03 Time Budget 04 Provider DNA 05 Simulation 06 Collusion 07 Invoice Testing 08 Rule Engine
01

Network Graph Analysis

Closed-Loop Detection

Traditional fraud detection examines individual claims in isolation. Network Graph Analysis maps the entire NDIS ecosystem as a connected graph -- providers, participants, workers, locations, and the relationships between them. This reveals hidden patterns that are completely invisible at the transaction level.

550+
Graph Nodes
10K+
Relationship Edges
5
Entity Types
Real-time
Graph Updates

What It Detects

  • Closed-loop money flows between connected providers
  • Invoice cycling through intermediary entities
  • Controlled provider clusters with shared infrastructure
  • Shared staff networks indicating coordinated operations
  • Shared addresses across multiple registered providers

How It Works

The engine builds a continuously updating directed graph using NetworkX. Nodes represent providers, participants, workers, and locations. Edges are weighted by billing volume, service relationships, and employment links. Cycle detection algorithms identify closed-loop flows, while community detection reveals suspicious clusters.

Real-World Scenario

Provider A pays Worker X. Worker X is also registered with Provider B. Provider B bills NDIS for services to Provider A's clients. The money flows in a circle -- this is invoice cycling, and it's invisible without graph analysis. CareIntegrity.AI detects this automatically by finding cycles in the provider-worker-participant graph.

Impact

In the demo dataset alone, the Network Graph engine identified 12 closed-loop alerts and multiple shared-staff clusters -- patterns that would take human auditors months to uncover through manual investigation.

02

Behavioural Drift Engine

Structural Impossibility

Fraudulent providers don't start fraudulent -- they drift. The Behavioural Drift Engine tracks each provider's "fingerprint" over time: billing patterns, session lengths, staffing ratios, geographic spread, and service mix. When a provider's behaviour changes faster than any legitimate business could, the engine flags it as structurally impossible.

12
Monthly Fingerprints
15+
Tracked Dimensions
222
Alerts Generated

What It Tracks

  • Average hours billed per participant per month
  • Session length distributions and variance
  • Participant growth rate vs workforce growth
  • Staffing ratio changes (participants per worker)
  • Geographic spread of service locations
  • Weekend and night billing ratios
  • Service mix evolution over time

Fraud Signals

  • Impossible acceleration: 5 to 80 participants in 6 weeks with no new staff
  • Billing spikes: 3x+ monthly billing increase without explanation
  • Staffing anomalies: 20:1+ participant-to-worker ratios
  • Time-of-day shifts: sudden move to after-hours billing
  • Geographic expansion without physical presence

Real-World Scenario

A small therapy provider with 5 participants and 2 workers suddenly grows to 80 participants in 6 weeks -- but still has the same 2 workers. This is a structural impossibility. No legitimate provider can serve 40 participants per worker. The Behavioural Drift Engine catches this by comparing growth trajectories against workforce capacity.

03

Time Budget Constraints

Physical Impossibility

Every human has exactly 24 hours in a day. Every worker can only be in one place at a time. Every journey between locations takes time. The Time Budget engine enforces these immutable physical laws against claimed service delivery, revealing fabricated services that violate the laws of physics.

100
Time Impossibilities
50
Excessive Hours
50
Travel Impossibilities
100
Overservicing Flags

Three Constraint Models

  • Worker simultaneity: Cannot bill at two locations at the same time
  • Daily capacity: No worker can deliver 24 hours of care in a day
  • Travel feasibility: Cannot travel 60km between sessions in 10 minutes

Participant Protection

  • Weekly hour allocation monitoring
  • Budget consumption tracking against plan limits
  • Service stacking detection (excessive concurrent therapies)
  • Peak week analysis vs allocated support hours

Real-World Scenario

Worker WRK-0005 is billed for a 3-hour session in Parramatta ending at 2:00 PM, and another 3-hour session in Campbelltown starting at 2:10 PM. These locations are 35km apart. Even at highway speed, this journey takes 30+ minutes. The Time Budget engine flags this as a travel impossibility -- one of these sessions is fabricated.

04

Provider DNA Embeddings

AI Representation Learning

Every provider has a "DNA" -- a unique pattern of how they bill, who they serve, what services they deliver, and when they work. The Provider DNA engine converts each provider into a high-dimensional vector embedding, then uses PCA and distance metrics to detect mutations, cluster anomalies, and providers that suddenly change their fundamental nature.

35
Embedding Dimensions
60
Providers Vectorised
30
Mutations Detected
170
Cluster Anomalies

Embedding Features

  • Service mix proportions (12 NDIS service types)
  • Participant demographics served (7 disability types)
  • Average session duration and rate
  • Weekend and night billing ratios
  • Geographic spread and billing intensity
  • Time distribution entropy (8 time-of-day bins)
  • Rate variance and session length variance

Detection Methods

  • Cosine distance between monthly embeddings (mutation detection)
  • Euclidean distance for magnitude changes
  • PCA projection to 2D with force-spread for visualisation
  • Pairwise similarity matrix against known fraud providers
  • Cluster proximity analysis using normalised vectors

Real-World Scenario

A registered Occupational Therapy provider's DNA suddenly shifts -- their embedding shows a dramatic move toward SIL (Supported Independent Living) billing patterns within 2 months. Their service mix goes from 90% therapy to 70% SIL. This is a semantic role shift anomaly -- legitimate providers don't fundamentally change what they do overnight. This often indicates the provider has been taken over or is exploiting a new billing category.

05

Synthetic Simulation

Baseline Comparison

What does "normal care" actually look like? The Synthetic Simulation engine answers this by generating realistic care baselines for each participant based on their support needs level, disability type, and plan allocation. It then compares actual billing against these baselines to detect over-servicing, inflated frequency, and unnecessary service stacking.

300
Participant Baselines
4
Support Levels
100
Anomalies Detected

Baseline Parameters

  • Expected weekly hours by support needs level
  • Expected sessions per week
  • Maximum concurrent service types
  • Expected therapy frequency
  • Weekend service ratio expectations
  • Weekly cost benchmarks

What It Catches

  • Over-servicing: actual hours 2x+ above expected baseline
  • Inflated frequency: 3x more sessions than clinically justified
  • Service stacking: OT + Psychology + Speech + Support Work all daily
  • Excessive therapy: therapy sessions far exceeding clinical norms

Real-World Scenario

A participant with "low" support needs (expected ~8 hours/week) is receiving 35 hours/week from a single provider -- 4x the expected baseline. They're simultaneously receiving Occupational Therapy, Psychology, Speech Therapy, and Support Work every day. The Synthetic Simulation engine flags this as service stacking -- there's no clinical reason for this intensity of concurrent services at a low support level.

06

Collusion Detection

Organised Fraud

The most damaging NDIS fraud isn't individual -- it's organised. Provider cartels coordinate to maximise billing through shared staff, shared addresses, and circular referral patterns. The Collusion Detection engine uses graph community detection algorithms to identify these hidden networks, rendered as an interactive 3D collusion map.

60
Providers Analysed
880
Provider-to-Provider Links
3D
Rotating Visualisation

Collusion Signals

  • Shared staff: workers registered across multiple providers
  • Shared participants: same clients billed by connected providers
  • Shared addresses: multiple providers at one physical location
  • Referral loops: participants bouncing between the same 2-3 providers
  • Common phone numbers, ABN structures, or ownership

Detection Method

The engine builds a weighted provider affinity graph. Edge weights combine shared staff (3x weight), shared locations (2x weight), and shared participants (1x weight). Greedy modularity community detection identifies tightly-connected clusters. High-density clusters with multiple shared resources are flagged as potential cartels.

Real-World Scenario

Four providers (PRV-0003, PRV-0007, PRV-0009, PRV-0011) share 8 workers between them, operate from 3 common addresses, and bill the same 45 participants. The Collusion Detection engine identifies this as a provider cartel with 0.62 network density -- these providers are operationally the same entity hiding behind multiple registrations to maximise billing.

07

Invoice Pressure Testing

Forensic Z-Score Analysis

Every single invoice is scored against multiple baselines simultaneously -- the provider's own history, the participant's pattern, peer group averages, workforce constraints, and geographic feasibility. The result is a composite fraud likelihood score that combines statistical deviation, network risk, and behavioural drift into a single actionable number.

98K+
Invoices Tested
3
Scoring Dimensions
84%
Highest Score

Scoring Formula

Fraud Likelihood = Deviation(0.4) x Network Risk(0.3) x Behavioural Drift(0.3)

  • Deviation: z-scores against provider, participant, and global baselines
  • Network Risk: provider's position in the ecosystem graph
  • Behavioural Drift: how much the provider has changed recently

Officer Actions

  • Forensic drill-down with z-score breakdown per dimension
  • Provider and participant context panels
  • One-click reject, approve, flag, or issue penalty
  • Batch analysis scheduling (5/10/15/30/60 minute intervals)
  • Configurable score thresholds for auto-flagging

Real-World Scenario

Invoice CLM-45892 charges $847 for a 9.5-hour session at $89/hour. The provider's average is 3.2 hours at $62/hour (hours z-score: +2.8 sigma). The participant normally receives 2-hour sessions (participant z-score: +3.1 sigma). The provider is also flagged by the Network Graph as part of a shared-staff cluster. Combined fraud likelihood: 78%. The fraud officer reviews the forensic evidence and issues a penalty with one click.

08

Custom Rule Engine

Configurable Rules

While the 7 AI engines detect patterns automatically, experienced fraud officers often know specific red flags unique to their region, provider type, or current investigation. The Custom Rule Engine lets officers define their own detection rules with multiple conditions, operators, and AND/OR logic -- turning institutional knowledge into automated detection.

7
Default Rules
10
Available Fields
16.8K
Matches Found
7
Operators

Rule Capabilities

  • Multiple conditions per rule (unlimited)
  • AND/OR logic between conditions
  • Fields: hours, rate, amount, start hour, weekend flag, service type, provider/worker ID
  • Operators: greater than, less than, equals, not equals, contains
  • Severity levels: critical, high, medium, low
  • Toggle rules on/off with instant re-evaluation

Default Rules Included

  • High hourly rate (above $95/hour)
  • Excessive session duration (over 8 hours)
  • After-hours billing (midnight to 5 AM)
  • High value claims (over $800 per invoice)
  • Weekend + long hours (weekend claims over 6 hours)
  • Late night high-value (after 10 PM, over $500)
  • Suspiciously round hours with high amounts

Real-World Scenario

A fraud officer investigating SIL providers notices a pattern: providers billing exactly 8.0 hours at exactly $70/hour every day, across multiple participants. She creates a custom rule: hours == 8.0 AND rate == 70.0 AND service_type contains "SIL". The rule immediately finds 340 matching claims across 3 providers -- confirming a coordinated billing template being used to generate fraudulent invoices.

See All 8 Engines in Action

Live demo processing 98,966 claims across 60 providers in real-time.

Access Live Demo
Login: admin / NDISAdmin2025!