VIGÍA ML runs four live, on-device models in the browser: an LSTM forecaster, an autoencoder anomaly detector, a neural fault classifier, and a recommendation layer that converts findings into actionable guidance.

The project embodies a broader philosophy of verifiable, self-hosted AI tooling, echoing the authors’ Universal Trust Adapter approach.

Key benefits include privacy-by-architecture, offline operation, zero marginal training/inference costs at the edge, and auditable transparency through visible model weights in DevTools.

The tech stack is React 18, TypeScript 5.7, TensorFlow.js 4.22, Vite 6, and Tailwind 4, with 140 tests; all data loading, training, evaluation, and export happen client-side under an AL-1.0 source-available license.

A live demo and source code are available: see the demo at alicelabs-llc.github.io/vigia-ml and the repository at github.com/alicelabs-llc/vigia-ml.

VIGÍA ML is a browser-based predictive monitoring console for gas wells, enabling production forecasting, anomaly detection, fault diagnosis, and operational recommendations entirely on-device without sending data to servers.

The project emphasizes data privacy and governance by ensuring all data remains on the local device, functioning offline as a Progressive Web App with zero inference cost.

The UI is Spanish-first to serve LATAM users, while the codebase remains standard TypeScript/React and the interface language is configurable.

Practically, the system detects slow drifts and early anomalies that fixed-threshold methods miss, and it provides interpretable reasoning for every recommendation rather than opaque outputs.

All four on-device models can be trained on user data, with visible training progress (loss curves) to build trust in model convergence.