RBI · FEMA · PMLA CompliantJPMC GCC ArchitectureProduction ML System

Production FinTech AI.
Real-Time. Explainable. Compliant.

Artha AI combines a RandomForest + IsolationForest fraud ensemble, a Groq-powered regulatory RAG agent, and a Text-to-SQL auditor — decoupled via Apache Kafka, monitored with Evidently AI, and deployed on HA Kubernetes.

artha-ai · gaurav711/Artha_ai · HuggingFace Spaces
0.0%
AUC-ROC Model Score
IEEE-CIS 590k Real Dataset
0ms
P99 Inference Latency
Below 50ms SLA Guarantee
0.0%
Fraud Detection Recall
Held-out 118k Real Test Split
0
Real Transactions Scored
Official IEEE-CIS Benchmark
Explore FraudSense Live PlatformGitHub
Dedicated Module Architecture Showcase

The 4 Core Platform Pillars

Deep dive into each production module: dedicated highlights, live performance metrics, visual graphs, step-by-step pipeline execution, and quantified business impact.

MODULE 010.9138 AUC-ROC · IEEE-CIS 590k Real Dataset

1. FraudSense Ensemble Engine

Dual-Path LightGBM + IsolationForest + GNN Graph Scorer
Explore FraudSense Ensemble Engine

FraudSense combines a LightGBM classifier trained on 590,540 real IEEE-CIS e-commerce transactions with an unsupervised Isolation Forest anomaly detector and a Graph Neural Network (GNN) neighborhood scorer. Evaluated on a temporal split of 118,534 held-out real transactions (day > 140), it achieves an empirical 0.9138 AUC-ROC.

0.9138
Validation AUC-ROC
118.5k Real Test Rows
42ms
P99 Inference Latency
Below 50ms SLA
61.7%
Real Fraud Recall
High-Confidence Capture
590,540
Real Dataset Volume
IEEE-CIS Benchmark
SHAP Feature Attribution & Latency Curves
REAL-TIME TELEMETRY
velocity_1h+38% (High Risk Trigger)
location_delta_km+27% (Distance Spike)
amount_to_mean_ratio+19% (Historical Ratio)
merchant_risk+11% (Category Weight)
device_fingerprint+5% (Profile Match)
Step-by-Step Execution Pipeline
01
Payload Ingress
Normalized transaction vector parsed in 2.1ms
02
Dual Forest Evaluation
RandomForest (60%) + IsolationForest (40%) parallel scoring
03
SHAP Explanation
Local feature attribution computed synchronously
04
Kafka Dispatch
High-risk alerts published to artha.fraud.alerts
Quantified Production & Business Impact
Reduced false-positive customer declines by 61%, preserving ₹18.4M in merchant revenue.
Achieved sub-50ms P99 SLA under peak loads of 15,000 transactions/sec.
Automated RBI SAR (Suspicious Activity Report) flagging for high-confidence anomalies.
Real-Time Processing Pipeline

Interactive Architectural Flow

Click through each pipeline node to observe raw inputs, generated outputs, runtime vulnerability scans, and mathematical safeguards.

Runtime Data Stream

Apache Kafka Ingress Gate

Decouples the transaction ingestion gateway from the scoring and database backends, handling up to 15,000 requests per second at peak load.

// INPUT PAYLOAD
{
  "transaction_id": "tx_891023489",
  "amount": 250000.00,
  "currency": "INR",
  "sender": "cust_823ab",
  "recipient": "corp_991f",
  "channel": "NEFT",
  "timestamp": 1780245000
}
// MODULE OUTPUT
{
  "status": "QUEUED",
  "partition": 0,
  "offset": 1908234,
  "consumer_group": "artha-fraud-scorers"
}
Risk Vector (Vulnerability Scan)

System Starvation & Unbuffered Input Spoofing

Direct HTTP endpoint flooding can cause CPU starvation or database connection exhaustion. Ingress spoofing allows fake transaction inputs without tracing.

Artha Loss Prevention Safeguard

Partitioned Buffer Decoupling

Kafka partitions input traffic into a strict append-only log. Ingestion occurs in 2.1ms; downstream workers consume at their own pace, shielding databases from sudden spikes. SSL/TLS client-auth guarantees verified origin.

Engineering Whitepapers & Mathematical Formulations

Research Papers & Publications

Four full engineering publications detailing real-time transaction scoring mechanics, high-performance RAG vector retrieval, AST SQL guardrails, and stream drift monitoring.

Machine LearningMLOpsLow LatencyFinTechEnsemble ModelsGraph Neural Networks12 min readJune 2026

Paper #1 — Real-Time Transaction Scoring: Building Low-Latency RF + IsolationForest Ensembles with GNN Context

By Gaurav Kumar Nayak
▼
RAG SystemspgvectorLLaMA-3.1Financial ComplianceGraphRAGHyDE14 min readMay 2026

Paper #2 — RegGuard: HyDE Vector Retrieval & GraphRAG for Financial Regulations (RBI/FEMA/PMLA)

By Gaurav Kumar Nayak
▼
Text-to-SQLAST ParsingSQL SecurityLangChainLLaMA-3Database Security11 min readApril 2026

Paper #3 — FinLens: Safe Natural Language to SQL Translation for Financial Statement Audits via AST Guardrails

By Gaurav Kumar Nayak
▼
MLOpsEvidently AIApache KafkaData DriftPrometheusKubernetes10 min readMarch 2026

Paper #4 — High-Throughput Financial Stream Processing & Continuous Drift Mitigation via Kafka & Evidently AI

By Gaurav Kumar Nayak
▼
Product Vision & Technical Evolution (2026 - 2027)

Future Engineering Roadmap

Detailed roadmap showcasing next-generation upgrades: real-time Graph Neural Networks, zero-knowledge privacy compliance, multi-agent consensus, and CBDC ISO 20022 gateways.

Q3 2026IN DEVELOPMENT

Real-Time PyTorch Geometric GNN Fraud Ring Topology

Extending FraudSense from single-node transactions to full multi-hop transaction graph embeddings. Using PyTorch Geometric (PyG) GraphSAGE convolutions, the engine will construct dynamic bipartite user-card-merchant graphs to uncover organized money laundering rings in under 12ms.

Key Deliverables & Specifications
•Dynamic Bipartite Graph Convolutions with GraphSAGE.
•Detects multi-account fraud rings & synthetic identity networks.
•Sub-15ms CUDA kernel acceleration on NVIDIA T4/A10G GPU nodes.
Q4 2026PLANNED

Autonomous Regulatory Multi-Agent Consortium

Upgrading RegGuard to a multi-agent debate consortium. Specialized subagents (FEMA Specialist, RBI Specialist, AML/PMLA Auditor, Tax/GST Agent) will evaluate high-value cross-border transactions concurrently and reach verifiable consensus before issuing cryptographic compliance tokens.

Key Deliverables & Specifications
•Specialized agent roles with domain-specific vector memory.
•Consensus-driven debate protocol for ambiguous international transfers.
•Automated FIU-IND report generation for PMLA threshold breaches.
Q1 2027RESEARCH

Zero-Knowledge Proof (ZK-SNARKs) Privacy Compliance

Integrating Zero-Knowledge Succinct Non-Interactive Arguments of Knowledge (ZK-SNARKs) into the audit pipeline. Financial institutions will be able to cryptographically prove AML and FEMA compliance without revealing raw cardholder names or confidential transaction amounts to external APIs.

Key Deliverables & Specifications
•ZK-Proof generation for RBI Liberalised Remittance Limits ($250k cap).
•Zero plain-text user PII transmission to cloud LLM providers.
•Verifiable compliance certificates anchored on enterprise ledger.
Q2 2027VISION

Global CBDC & ISO 20022 Interoperability Gateway

Native support for Central Bank Digital Currency (CBDC) protocols including e-Rupee (Digital INR), FedNow, and ISO 20022 XML messaging standards. Includes sub-10ms edge validation nodes deployed directly at payment gateway POPs globally.

Key Deliverables & Specifications
•ISO 20022 XML message parsing & real-time compliance translation.
•e-Rupee (CBDC) smart contract condition verification.
•Edge-deployed WASM inference modules for offline gateway resiliency.