RustKernels

High-performance GPU kernel library for financial services, compliance, and enterprise analytics.

---

Why RustKernels?

Financial institutions face a common challenge: implementing high-performance analytics that scale from batch processing to real-time streaming while maintaining regulatory compliance. RustKernels solves this by providing:

• Battle-tested algorithms ported from production C# Orleans grains
• Nanosecond-scale latency for real-time fraud detection and order matching
• Unified API across 14 specialized domains
• GPU acceleration with automatic CPU fallback

// Detect AML patterns in transaction graphs - sub-millisecond response
let detector = CircularFlowRatio::new();
let risk_score = detector.compute(&transaction_graph)?;
if risk_score > 0.8 {
    alert_compliance_team(&transaction);
}

---

Overview

RustKernels provides 106 GPU-accelerated algorithms across 14 domain-specific crates, purpose-built for financial institutions, compliance teams, and enterprise analytics platforms.

What Makes RustKernels Different

Challenge	RustKernels Solution
Latency requirements vary widely	Dual execution modes: Batch (10-50μs) or Ring (100-500ns)
Complex multi-kernel workflows	Built-in K2K coordination patterns
Production reliability concerns	Ported from battle-tested C# implementations
GPU availability uncertainty	Automatic CPU fallback when CUDA unavailable
Regulatory explainability	SHAP values and feature importance kernels

---

Performance Characteristics

Metric	Batch Mode	Ring Mode
Launch overhead	10-50μs	N/A (persistent)
Message latency	N/A	100-500ns
State location	CPU → GPU transfer	GPU-resident
Throughput (PageRank)	~100K nodes/sec	~500K updates/sec
Memory efficiency	Standard	Optimized (persistent)

When to Use Each Mode

Batch Mode - Best for scheduled, heavy computation:

• End-of-day risk aggregation
• Batch AML screening (millions of transactions)
• Monthly compliance reporting
• Model training and backtesting

Ring Mode - Best for real-time, high-frequency operations:

• Order book matching (sub-millisecond)
• Real-time fraud scoring
• Streaming anomaly detection
• Live transaction monitoring

---

Domain Coverage

Domain	Crate	Kernels	Key Algorithms
Graph Analytics	rustkernel-graph	28	PageRank, Louvain, GNN inference, graph attention, cycle detection
Statistical ML	rustkernel-ml	17	K-Means, DBSCAN, isolation forest, federated learning, SHAP
Compliance	rustkernel-compliance	11	Circular flow detection, rapid movement, sanctions screening
Temporal Analysis	rustkernel-temporal	7	ARIMA, Prophet decomposition, change point detection
Risk Analytics	rustkernel-risk	5	Monte Carlo VaR, credit scoring, stress testing
Process Intelligence	rustkernel-procint	7	DFG construction, conformance checking, digital twin
Behavioral Analytics	rustkernel-behavioral	6	Profiling, forensic queries, causal analysis
Treasury	rustkernel-treasury	5	Liquidity optimization, FX hedging, NSFR calculation
Clearing	rustkernel-clearing	5	Multilateral netting, DVP matching, settlement
Accounting	rustkernel-accounting	9	Network generation, GL reconciliation, GAAP detection
Banking	rustkernel-banking	1	Fraud pattern matching (Aho-Corasick)
Order Matching	rustkernel-orderbook	1	High-frequency order book engine
Payments	rustkernel-payments	2	Payment processing, flow analysis
Audit	rustkernel-audit	2	Feature extraction, hypergraph construction

---

Use Cases

Anti-Money Laundering (AML)

Detect layering, structuring, and circular transaction patterns:

use rustkernel::graph::cycles::ShortCycleParticipation;
use rustkernel::compliance::circular_flow::CircularFlowRatio;

// Detect nodes participating in suspicious cycles
let cycle_detector = ShortCycleParticipation::new();
let results = cycle_detector.compute_all(&transaction_graph);

for result in results.iter().filter(|r| r.risk_level == CycleRiskLevel::Critical) {
    // Nodes in 4-cycles are high-priority for investigation
    flag_for_investigation(result.node_index);
}

// Compute circular flow ratios
let cfr = CircularFlowRatio::new();
let scores = cfr.compute_batch(&graph);

Real-Time Fraud Detection

Score transactions in real-time with streaming anomaly detection:

use rustkernel::ml::streaming::StreamingIsolationForest;

let detector = StreamingIsolationForest::new(StreamingConfig {
    num_trees: 100,
    sample_size: 256,
    window_size: 10000,
});

// Process incoming transactions
for transaction in transaction_stream {
    let score = detector.score(&transaction.features)?;
    if score > 0.7 {
        block_transaction(&transaction);
    }
}

Process Mining & Digital Twin

Simulate process changes before deployment:

use rustkernel::procint::simulation::{DigitalTwin, SimulationConfig};

let twin = DigitalTwin::new();

// Simulate adding 2 more resources to bottleneck activity
let what_if = twin.simulate(&process_model, &SimulationConfig {
    num_simulations: 1000,
    resource_overrides: vec![("Review", 5)], // 3 → 5 reviewers
    ..Default::default()
})?;

println!("Projected improvement: {:.1}% faster",
    (1.0 - what_if.avg_completion_time / baseline.avg_completion_time) * 100.0);

Graph Neural Networks for Entity Resolution

Link prediction and entity matching with GNN inference:

use rustkernel::graph::gnn::{GNNInference, GNNConfig};

let gnn = GNNInference::new();

let embeddings = gnn.infer(&entity_graph, &node_features, &GNNConfig {
    hidden_dim: 64,
    num_layers: 2,
    aggregation: AggregationType::Mean,
})?;

// Find similar entities via embedding similarity
let matches = find_similar_embeddings(&embeddings, threshold: 0.9);

---

Recent Additions

The latest release introduces innovative kernel categories:

Category	Kernels	Description
Graph Neural Networks	GNNInference, GraphAttention	Message-passing GNN and multi-head attention for node classification
NLP/Embeddings	EmbeddingGeneration, SemanticSimilarity	TF-IDF embeddings and document similarity
Federated Learning	SecureAggregation	Privacy-preserving model aggregation with differential privacy
Healthcare Analytics	DrugInteractionPrediction, ClinicalPathwayConformance	Clinical decision support kernels
Process Simulation	DigitalTwin	Monte Carlo process simulation and what-if analysis
Streaming ML	StreamingIsolationForest, AdaptiveThreshold	Online anomaly detection with concept drift handling
Explainability	SHAPValues, FeatureImportance	Model interpretability for regulatory compliance

---

Installation

Add RustKernels to your Cargo.toml:

[dependencies]
rustkernel = "0.1.0"

Feature Flags

Control which domains are compiled to optimize binary size:

# Default features (graph, ml, compliance, temporal, risk)
rustkernel = "0.1.0"

# Selective domain inclusion
rustkernel = { version = "0.1.0", features = ["graph", "compliance", "procint"] }

# All domains
rustkernel = { version = "0.1.0", features = ["full"] }

---

Quick Start

use rustkernel::prelude::*;
use rustkernel::graph::centrality::PageRank;

#[tokio::main]
async fn main() -> Result<()> {
    // Create kernel instance
    let kernel = PageRank::new();

    // Access metadata
    let metadata = kernel.metadata();
    println!("Kernel: {} ({})", metadata.id, metadata.domain);

    // Build input
    let input = PageRankInput {
        num_nodes: 1000,
        edges: load_edges()?,
        damping_factor: 0.85,
        max_iterations: 100,
        tolerance: 1e-6,
    };

    // Execute
    let result = kernel.execute(input).await?;

    println!("Converged in {} iterations", result.iterations);
    println!("Top node: {} (score: {:.4})",
        result.top_node(),
        result.scores[result.top_node()]);

    Ok(())
}

---

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                         rustkernel                               │
│                    (facade crate, re-exports)                    │
├─────────────────────────────────────────────────────────────────┤
│  rustkernel-core          │  rustkernel-derive                  │
│  - GpuKernel trait        │  - #[gpu_kernel] macro              │
│  - BatchKernel trait      │  - #[derive(RingMessage)]           │
│  - RingKernelHandler      │                                     │
│  - K2K coordination       │                                     │
│  - License validation     │                                     │
├─────────────────────────────────────────────────────────────────┤
│                    Domain Crates (14)                            │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐   │
│  │  graph  │ │   ml    │ │complianc│ │temporal │ │  risk   │   │
│  │  (28)   │ │  (17)   │ │  (11)   │ │   (7)   │ │   (5)   │   │
│  └─────────┘ └─────────┘ └─────────┘ └─────────┘ └─────────┘   │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐   │
│  │procint  │ │behavior │ │treasury │ │clearing │ │accounting│   │
│  │   (7)   │ │   (6)   │ │   (5)   │ │   (5)   │ │   (9)   │   │
│  └─────────┘ └─────────┘ └─────────┘ └─────────┘ └─────────┘   │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐               │
│  │ banking │ │orderbook│ │payments │ │  audit  │               │
│  │   (1)   │ │   (1)   │ │   (2)   │ │   (2)   │               │
│  └─────────┘ └─────────┘ └─────────┘ └─────────┘               │
├─────────────────────────────────────────────────────────────────┤
│                    RustCompute / RingKernel                      │
│                    (GPU execution framework)                     │
└─────────────────────────────────────────────────────────────────┘

---

Requirements

Requirement	Version	Notes
Rust	1.85+	Edition 2024 features required
RustCompute	Latest	RingKernel framework (path dependency)
CUDA Toolkit	12.0+	Optional; falls back to CPU if unavailable

---

Building and Testing

# Build entire workspace
cargo build --workspace

# Run all tests
cargo test --workspace

# Test specific domain
cargo test --package rustkernel-graph
cargo test --package rustkernel-ml

# Run benchmarks
cargo bench --package rustkernel

# Generate documentation
cargo doc --workspace --no-deps --open

# Lint
cargo clippy --all-targets --all-features -- -D warnings

---

Documentation

• Online Documentation - Comprehensive guides and API reference
• Kernel Catalogue - Complete listing of all 106 kernels
• Architecture Guide - System design and patterns

---

Contributing

Contributions are welcome. Please see CONTRIBUTING.md for guidelines.

---

License

Licensed under the Apache License, Version 2.0. See LICENSE for details.

---

Author: Michael Ivertowski Version: 0.1.0 Kernels: 106 across 14 domains