Find known and unknown anomalies in your data through geometric measurement.
GDI is Geometric Deterministic Intelligence: anomaly detection by measuring where points sit in mathematical space -- not by fitting another statistical or neural model to your problem.
Normal behavior clusters. Anomalies sit away from that cluster. The distance and angle from a baseline are facts about the data, not estimates. Same input, same output, every run.
14 public datasets, identical code, zero per-dataset configuration, CPU-only -- completed back-to-back in 156 seconds. The geometry was always in the data. GDI measures it.
Anomaly detection through measurement
GDI does not watch an AI system from the outside. It analyzes the vectors your data already forms -- and surfaces both the rare events you expect and the novel deviations you have not seen before.
The anomalies you know matter but are hard to catch: fraud at fractions of a percent, intrusions buried in normal traffic, rare failures in industrial telemetry. GDI refines detection when events are rare so alerts stay actionable.
Patterns you have never labeled and cannot sign in advance. Because detection is geometric -- not a fixed classifier -- novel points that sit far from the established baseline surface without retraining or new rules.
Most systems fit millions of parameters and approximate patterns. GDI measures distances and angles in the space your data already occupies -- like using a ruler, not training another model.
When a point is flagged, the reason is the measurement itself: how far it sits from the baseline, and which features drove that distance. No separate interpretability layer.
Raw numerical telemetry or embeddings from another system. You send data; GDI selects the right geometric measurement for its dimensionality and shape.
Anomaly scores, drift, and attributions are themselves vectors. The same engine can analyze those outputs to reveal fleet-wide or enterprise-wide patterns no single stream shows alone.
Three common deployment paths
Run geometric detection on telemetry, transactions, sensors, or logs expressed as vectors. Standalone product -- no Atum required.
Embeddings translate complex objects into geometry; GDI measures that geometry. Useful for LLM output drift, hallucination signals, and any domain where an embedding model already exists.
Neural systems generate and reason; geometric systems detect, classify, and explain. Use both when you need provable models and continuous measurement on the same streams.
GDI does not replace generative or reasoning systems. Embeddings can translate a domain into geometry; geometric measurement can flag drift, degradation, and hallucination in those outputs. Triaxis Atum extends that partnership when a workload needs deterministic, provable inference underneath.
Explore Triaxis AtumHow teams apply geometric detection
The same measurement engine supports operational screening, population analytics, and compliance review -- without separate products or domain-specific models.
Operations
Every vector in the stream is measured against the baseline -- not a sampled subset of the traffic.
Each flag reports the measurements behind it: which features deviated, by how much, and from what baseline.
Sudden spikes, slow drift, and structural shifts can surface in parallel instead of fighting over one threshold.
Analytics
See which features drive anomalies across a population, not just point by point.
Track how the geometry of your data shifts over time while the baseline adapts to legitimate slow change.
Feed anomaly scores from many sources back into the same engine to find correlations invisible locally.
Compliance
Single Docker container, no GPU, no external dependencies -- runs inside your perimeter.
Every detection is a reproducible geometric fact about the data, not an opaque model score.
Same input produces the same output -- suitable for regulated review and replay.
Named datasets, measured F1 scores
Performance is dataset-dependent -- so we publish the datasets, not a vague range. GDI is unsupervised on every run below: no training labels, no per-dataset retuning.
14 public datasets, identical code, zero per-dataset configuration, CPU-only -- completed back-to-back in 156 seconds.
Near-supervised F1 on CLAP embeddings with no training labels.
Matches supervised deep learning without labeled fault data.
96.3% recall on hazardous air events at very high throughput.
Peak throughput benchmark -- one CPU, massive meter fleets.
Near-perfect precision for SOC workflows on behavioral embeddings.
Instability detection for grid modernization workloads.
| Dataset | Domain | Dims | F1 | Precision | Throughput | Context |
|---|---|---|---|---|---|---|
| ESC-50 Audio Anomaly | Environmental sound / security audio | 512D | 97.6% | 97.3% | 1,242/sec | Top performer |
| Tennessee Eastman Process | Industrial chemical fault detection | 52D | 93.0% | 98.3% | 6,239/sec | Top performer |
| Beijing Air Quality | Environmental / smart city sensors | 12D | 90.8% | 86.0% | 67,299/sec | Top performer |
| Household Power Consumption | Smart grid / IoT telemetry | 7D | 90.6% | 99.2% | 59,515/sec | Top performer |
| CERT Insider Threat | User behavior analytics | 768D | 89.7% | 100.0% | 1,051/sec | Competitive |
| HaluEval LLM Hallucination | LLM output monitoring | 384D | 83.2% | 89.4% | 2,359/sec | Top performer |
| FI-2010 High-Frequency Trading | Market microstructure | 40D | 80.4% | 99.1% | 27,859/sec | Competitive |
| Elliptic Bitcoin AML | Blockchain compliance | 166D | 80.9% | 75.3% | 4,390/sec | Competitive |
| Smart Grid Stability | Power grid simulation | 12D | 84.0% | 77.0% | 16,388/sec | Competitive |
| NSL-KDD Network Intrusion | Explainable network IDS | 41D | 78.6% | 85.5% | 22,622/sec | Strong |
| NASA Turbofan Engine | Predictive maintenance | 24D | 78.6% | 68.4% | 16,889/sec | Strong |
| CICIDS-2017 Network Attacks | Modern network security | 76D | 79.4% | 67.3% | 9,085/sec | Strong |
| CTU-13 Botnet Detection | Network PCAP / intrusion | 384D | 72.0% | 69.6% | 1,713/sec | Strong |
| Credit Card Fraud | Extreme imbalance fraud | 29D | 36.0% | 40.0% | 36,551/sec | Dataset-dependent |
Hard datasets stay in the table on purpose. Credit card fraud at 36% F1 is still the best unsupervised result on that benchmark -- useful as a fast screen, not a claim of universal 90%+ F1. Ask us for the domain closest to your data.
See geometric detection on your own data.
Send vectors. Get measurements. No labeled training data, no GPU, and no domain-specific model to rebuild for each new source.