gpartin/WaveGuardClient

Anomaly detection powered by wave physics. Not machine learning.
One API call. Fully stateless. Works on any data type.
Benchmarks •
Quickstart •
Use Cases •
Examples •
MCP / Claude •
API Reference
---
What is WaveGuard?
WaveGuard is a **general-purpose anomaly detection API**. Send it any data — server metrics, financial transactions, log files, sensor readings, time series — and get back anomaly scores, confidence levels, and explanations of *which features* triggered the alert.
**No training pipelines. No model management. No state. One API call.**
Under the hood, it uses GPU-accelerated wave physics instead of machine learning. You don't need to know or care about the physics — it's all server-side.
Modal dashboard vs API endpoints
If you look at Modal, you will see deployed **functions** (for example fastapi_app, gpu_scan, gpu_fingerprint).
Those are compute/runtime units, not the HTTP route list.
To see all live API endpoints, use:
- OpenAPI docs: https://gpartin--waveguard-api-fastapi-app.modal.run/docs
- OpenAPI JSON: https://gpartin--waveguard-api-fastapi-app.modal.run/openapi.json
How does it actually work?
Your data is encoded onto a 64³ lattice and run through coupled wave equation simulations on GPU. Normal data produces stable wave patterns; anomalies produce divergent ones. A 52-dimensional statistical fingerprint is compared between training and test data. Everything is torn down after each call — nothing is stored.
The key advantage over ML: no training data requirements (2+ samples is enough), no model drift, no retraining, no hyperparameter tuning. Same API call works on structured data, text, numbers, and time series.
Benchmarks (v2.2)
**WaveGuard v2.2 vs scikit-learn** across 6 real-world scenarios (10 training + 10 test samples each).
> **TL;DR**: WaveGuard v2.2 **wins 4 of 6 scenarios** and averages 0.76 F1 — competitive with sklearn methods while requiring zero ML expertise.
F1 Score (balanced precision-recall)
| Scenario | WaveGuard | IsolationForest | LOF | OneClassSVM |
|----------|:---------:|:---------------:|:---:|:-----------:|
| Server Metrics (IT Ops) | **0.87** | 0.71 | 0.87 | 0.62 |
| Financial Fraud | **0.83** | 0.74 | 0.77 | 0.77 |
| IoT Sensors (Industrial) | **0.87** | 0.69 | 0.69 | 0.65 |
| Network Traffic (Security) | **0.82** | 0.61 | 0.77 | 0.61 |
| Time-Series (Monitoring) | 0.46 | 0.77 | **0.80** | 0.67 |
| Sparse Features (Logs) | 0.72 | **0.90** | 0.82 | 0.78 |
| **Average** | **0.76** | 0.74 | **0.79** | 0.68 |
What's new in v2.2
Multi-resolution scoring tracks each feature's **local lattice energy** in addition to global fingerprint distance. This catches subtle per-feature anomalies (like 3 of 10 IoT sensors drifting) that v2.1's global averaging missed. IoT F1 improved from 0.30 → 0.87.
When to choose WaveGuard over sklearn
| Choose WaveGuard when... | Choose sklearn when... |
|--------------------------|------------------------|
| False alarms are expensive (alert fatigue, SRE pages) | You need to catch every possible anomaly |
| You have no ML expertise on the team | You have data scientists who can tune models |
| You need a zero-config API call | You can manage model lifecycle (train/save/load) |
| Data schema changes frequently | Feature engineering is stable |
| Your AI agent needs anomaly detection (MCP) | Everything runs locally, no API calls |
Reproduce these benchmarks
Results saved to benchmarks/benchmark_results.json. Benchmarks use deterministic random seeds for reproducibility.
> **Expanded benchmarks**: WaveGuard ranks #1 in F1 score on all 12 public benchmark datasets. See the full comparison on HuggingFace.
Real-World Validation: Crypto Crash Detection
WaveGuard powers CryptoGuard, a crypto risk scanner. Backtested against 7 historical crashes (LUNA, FTX, Celsius, 3AC, UST, SOL/FTX, TITAN):