QuirkBot.ai

AI-powered transaction analysis — 80 dimensions, 25 engines, complete explainability

---

What is QuirkBot.ai?

Financial fraud detection is either opaque (black-box ML models that auditors can't explain) or simplistic (rule-based systems that sophisticated fraud easily bypasses). Compliance teams need both intelligence and transparency — every decision must be traceable for SOX, GDPR, and AML requirements.

QuirkBot scores every transaction across 80 risk dimensions (40 inherent risk + 40 control risk) using 25 specialized AI engines. Results are projected into interactive 3D space via PCA. Every decision is fully explainable — click any data point for the complete dimensional breakdown, engine firing history, and audit trail.

The 25 AI Engines

Phase 1 — Inherent Risk Detection (11 engines)

#	Engine	What It Detects
1	Large Amount	Adaptive P95 threshold breaches
2	Off Hours	Transactions outside business hours
3	Velocity Spike	3-window burst patterns
4	Cross Border	Multi-field geography anomalies
5	Round Amount	Multi-base structuring (100s, 500s, 1000s)
6	Budget Overrun	Inferred budget threshold breaches
7	Dept Spike	Department-relative baseline deviations
8	Metric Deterioration	8-KPI cross-metric degradation
9	Benford's Law	Leading digit distribution analysis
10	Description Anomaly	Vague or suspicious descriptions
11	Amount Splitting	Fragmentation / structuring patterns

Phase 2 — Mixed Detection (5 engines)

#	Engine	What It Detects
12	High-Value Refund	Refund velocity + ratio anomalies
13	Weekend Posting	Saturday/Sunday journal entries
14	Revenue No Cash	Cash match window gaps
15	Cross-Statement	IS/CF/BS reconciliation failures
16	Journal Entry	Multi-signal journal red flags

Phase 3 — Control Risk (5 engines)

#	Engine	What It Detects
17	Unusual Pairing	Debit/credit frequency anomalies
18	Duplicate Ref	Exact + fuzzy + economic duplicates
19	Process Delay	SLA threshold breaches
20	Segregation of Duties	Incompatible role combinations
21	Control Weakness	Ultimate control risk assessment

Phase 4 — Mitigation (4 engines)

#	Engine	What It Does
M1	Verified Cleared	IR reduction for verified counterparties
M2	Historical Anchor	IR reduction for established patterns
M3	MFA Authenticated	CR reduction for strong auth evidence
M4	Reconciled Receipt	IR+CR reduction for documentary evidence

Risk Pipeline

Stage 1: INITIALIZE           → inherentRisk = 0.05, controlRisk = 0.00
Stage 2: DETECTION             → 21 engines fire sequentially
Stage 3: DERIVE CONTROL RISK   → Based on inherent risk + random variance
Stage 4: MITIGATION            → 4 engines reduce IR/CR (max 50% cap)
Stage 5: NET RISK + FLOOR      → Forensic floor 0.02 preserves audit trail

Decision Output

Decision	Risk Range	Action
ACCEPT	<20%	Process normally
LOW_RISK	20–40%	Process with logging
MONITOR	40–60%	Queue for review
REVIEW	60–80%	Escalate to investigator
DECLINE	80%+	Reject transaction

3D Visualization

• Three.js with InstancedMesh for 100K+ data points
• Client-side PCA/SVD reducing 80 dimensions to 3D space
• Multiple views: Geographic, metric-based, risk-based, time-series
• Full explainability: Click any point to see all 80 dimensions

Tech Stack

Layer	Technology
Backend	FastAPI (Python, async)
Database	PostgreSQL + SQLAlchemy Async + Alembic
Cache	Redis
Frontend	Vite + JavaScript + Three.js
Ingestion	CSV/JSON upload, max 50MB / 500K rows
Data Integrity	Immutable raw_json preservation

Who Is This For?

• Financial institutions detecting fraud before it clears
• Internal audit teams needing SOX/GDPR-ready explainability
• Compliance departments monitoring control environments
• AML/KYC teams tracking money laundering patterns
• CFOs wanting real-time risk visibility across the organization

What Makes It Different

Feature	QuirkBot	Traditional ML	Rule-Based
Explainability	Full 80-D breakdown	Black box	Limited
Adaptiveness	Self-learning baselines	Requires retraining	Manual updates
Dimensions	80 (40 IR + 40 CR)	Varies, opaque	<10 rules
Forensic trail	0.02 floor, never zero	Often zeros out	Binary pass/fail
Visualization	Real-time 3D	Dashboards	Tables

Getting Started

Visit quirkbot.ai to explore the platform. Free tier includes 5,000 data points.

About

Built by F² AI in South Africa. Deployed globally on Google Cloud Platform.

License

MIT