Triaxis AICore Engine
A detection engine designed to enhance, augment, or replace anomaly detection in your existing products. Bimodal architecture delivers orders of magnitude speed improvement on CPU alone.
Triaxis AI is not a standalone product or SaaS. You provide the data, you build the UI, you own the customer relationship. We provide the detection engine.
How the Triaxis AI Engine Works
A brief overview of our geometric signal detection approach
Bimodal Architecture
Phase 1: Learning
MST-Based Clustering
Uses the APD (Adaptive Parameter Discovery) algorithm to discover natural cluster boundaries through Torque-based MST analysis. Runs once during initial learning or when drift is detected.
Phase 2: Scoring
Linear-Time Detection
Lightning-fast O(N) anomaly detection using vectorized cosine similarity against learned prototypes. Enables real-time processing of high-velocity data streams.
Up to 400x Faster Than Re-Clustering
Performance at Scale
Real throughput on commodity hardware—no GPUs required.
Works on Any Data Type
Time series, text, embeddings, network traffic, sensor streams—Triaxis AI extracts signals from the geometry.
Enterprise Features
Hardware-Locked Licensing
RSA-signed JWT licenses with hardware fingerprinting. Supports air-gapped deployments with offline validation.
Drift Detection
Automatic drift detection via anomaly rate monitoring. Triggers Phase 1 re-learning when distribution shifts detected.
CPU-Only Processing
No GPU required. Vectorized operations achieve 84K+ samples/sec average throughput on standard CPUs.
AI You Can Actually Audit
Triaxis AI fuses three branches of artificial intelligence into a deterministic system - without the GPU requirements, training data, or black-box opacity of neural networks.
Unsupervised Clustering
Discovers natural groupings in data without labeled training examples
Geometric Analysis
Uses cosine similarity and high-dimensional geometry to detect behavioral deviations
Statistical Learning
Builds distribution models from observed data to score new observations
Reproducible
Run the same data twice, get identical results. No stochastic variation.
Auditable
Every detection includes feature-level explanations with z-scores and baseline statistics.
No Hallucinations
Geometric math cannot invent false patterns. Results are grounded in actual data.
Regulation-Ready
Meets explainability requirements for financial services, healthcare, and critical infrastructure.
Compliance-Ready Data Processor
Triaxis AI processes numerical vectors, not PII. We provide compliant infrastructure; you control what data flows through it.
No PII Storage by Design
Engine processes embeddings and feature vectors - raw data stays in your systems
Infrastructure Hooks for Compliance
Built-in support for data deletion, audit logging, and TTL policies
Licensee Responsibility Model
Similar to AWS/GCP - we provide compliant infrastructure, you ensure compliant data flows
Data Flow Architecture
Benchmark Results
14 diverse datasets, identical code, zero configuration changes. All benchmarks run on CPU-only hardware.
| Application | Dims | F1 Score | Precision | Throughput | Ranking |
|---|---|---|---|---|---|
Blockchain AML Detection Elliptic Bitcoin dataset | 166D | 80.9% | 79.3% | 31,184/sec | STRONG |
Industrial Fault Detection Tennessee Eastman Process | 52D | 93.0% | 98.3% | 34,887/sec | #1 UNSUP |
Smart Grid Monitoring Household Power Consumption | 7D | 90.6% | 99.2% | 259,643/sec | CHAMPION |
Insider Threat Detection CERT Insider Threat | 768D | 89.7% | 100.0% | 6,627/sec | #1 UNSUP |
LLM Hallucination Detection HaluEval Q/A Coherence | 384D | 83.2% | 89.4% | 18,057/sec | CHAMPION |
High-Frequency Trading FI-2010 Limit Order Book | 40D | 80.4% | 99.1% | 115,231/sec | COMPETITIVE |
Network Intrusion (NSL-KDD) DOS/Probe/R2L/U2R attacks | 41D | 78.6% | 85.5% | 186,715/sec | STRONG |
Audio Anomaly Detection ESC-50 Environmental Sounds | 512D | 97.6% | 97.3% | 5,228/sec | CHAMPION |
Network Intrusion (CTU-13) Botnet PCAP Analysis | 384D | 72.0% | 69.6% | 12,823/sec | STRONG |
Financial Fraud Detection Credit Card (0.17% fraud rate) | 29D | 36.0% | 40.0% | 176,713/sec | #1 UNSUP |
Environmental Monitoring Beijing Air Quality (PM2.5/O3/NO2) | 12D | 90.8% | 86.0% | 67,299/sec | CHAMPION |
Power Grid Stability Smart Grid Simulation (12D) | 12D | 84.0% | 77.0% | 16,388/sec | #1 UNSUP |
Modern Network Security CICIDS-2017 DoS/Patator attacks | 76D | 79.4% | 67.3% | 9,085/sec | #1 UNSUP |
Predictive Maintenance NASA Turbofan Engine (24 sensors) | 24D | 78.6% | 68.4% | 16,889/sec | #1 UNSUP |
Key Achievement: All 14 benchmarks run with identical code and zero configuration changes.
Total benchmark suite completes in 156 seconds on a standard CPU with no GPU.