Attach a monitor to a PyTorch model, calibrate on clean data, then score inference batches and training updates -- entirely in-process, no daemon, no network calls. Pre-built artifacts for 90+ HuggingFace models available in the registry.
The SDK works entirely offline out of the box. No service footprint, no weight transfer, no external calls -- just load an artifact and score.
Load `./integrity.json` and score in-process. Works in notebooks, CI pipelines, and production inference loops. Zero network dependency.
Deploy entirely inside a VPC or on-premise with no internet connectivity. Raw model weights never leave your environment.
The production layer is built around a very small operational loop: calibrate, save, load, score, and optionally upload the artifact to a service.
from signai import monitor
m = monitor.attach(model, num_classes=10, device="cuda")
m.calibrate(clean_loader, device="cuda", phase="inference", calib_batches=200)
m.save("./integrity.json")
m = monitor.load(model, artifact="./integrity.json", device="cuda")
result = m.score_inference(x, y)
Start free. Upgrade when you need nn and assoc detectors, history, and alerts. Enterprise contracts available for air-gapped and compliance deployments.
No account needed. Start monitoring immediately after install.
All three detectors, 200 calibrations/month, history and alerts.
Unlimited calibrations, air-gapped deployment, compliance and dedicated support.
License key delivered by email within 1 hour of payment. | All plans include a 3-day free trial of Pro features.
The repository now includes product-facing setup instructions, deployment notes, runtime examples, and troubleshooting guidance for the production layer.
Updated with install paths, deployment modes, CLI usage, Docker commands, privacy boundaries, and edition guidance.
A customer-facing walkthrough covering local mode, remote mode, calibration, runtime scoring, authentication, troubleshooting, and rollout guidance.
A focused operations doc for self-hosted, Docker, SQLite, and Postgres-backed setups, including health checks and production recommendations.
Release-facing docs now capture the production-layer delta and the final verification list for packaging, Docker, docs, and customer handoff.
Customer copy now explicitly states that only compact behavioral vectors are transmitted and that raw vectors are discarded after scoring.