SensorOps.
A maintenance workflow, end to end.

The question

What does it take to connect a maintenance model to a usable workflow? SensorOps explores the infrastructure around that question, using replayed sample data rather than live factory equipment.

What I built

The application connects CSV ingestion to feature preparation, model comparison, MLflow tracking and staging gates, and a FastAPI prediction service. The model experiments include Isolation Forest, an LSTM autoencoder, and an echo-state network.

An optional assistant uses ChromaDB retrieval and a DeepSeek-backed interface to help explore documentation and explanations. The code also includes a SHA-256 hash-chained logging component for recording events.

The engineering focus

• Orchestration: expressing ingestion and modelling steps as a Dagster workflow.
• Experiment tracking: keeping model runs and staging decisions visible through MLflow.
• Application boundaries: exposing prediction through a typed API instead of a notebook-only interface.
• Operational context: bringing retrieval and event records alongside model output.

What this prototype does — and does not establish

SensorOps demonstrates how these components can fit together. Its demonstration data includes synthetic vibration generated using known failure labels, so it is not suitable evidence of predictive accuracy or generalization to a factory.

This is a portfolio prototype, not a validated industrial deployment. The logging component is a technical feature, not a claim of regulatory compliance.