Predictive Maintenance Platform
A platform that watches industrial equipment sensors in real time and warns maintenance teams before a machine breaks down, instead of after.
The problem
Industrial equipment maintenance is usually reactive (fix it after it breaks) or scheduled on a fixed calendar regardless of actual condition. Both waste money: unplanned downtime is expensive, and scheduled maintenance replaces parts that still had useful life left. Predicting failures from real-time sensor data needed an architecture that could ingest high-volume streams from many tenants without one tenant's load affecting another's.
The approach
Built an event-driven platform: equipment sensors publish over MQTT, which feeds into Kafka for durable, multi-tenant stream processing; Python and FastAPI services train and serve the failure-prediction models, with model versioning and a human-in-the-loop review step before a retrained model goes live; a Next.js and React dashboard surfaces equipment health and notifications, backed by PostgreSQL and Redis, with secure authentication throughout. My role was the ingestion service: the Go component that takes in the raw MQTT sensor streams and reliably hands them off into Kafka for every tenant.
What I learned
Building the ingestion service meant designing for backpressure and tenant isolation from the start: a burst of sensor data from one client's equipment couldn't be allowed to slow down or drop data for anyone else. Go's concurrency model made the high-throughput, low-latency ingestion path much more manageable than it would have been otherwise.