Meter telemetry at scale feeding demand forecasts and asset-failure prediction, with figures traceable to source readings.
Where it started
Meter and sensor readings arrived late, out of order and occasionally twice, which made every downstream figure arguable. Demand forecasting ran overnight on a subset of the data, and asset maintenance was scheduled on a fixed interval regardless of condition.
- 01
Built a streaming ingestion layer tolerant of late and duplicate readings, with watermarking and deterministic reprocessing.
- 02
Modelled demand with weather and price covariates, publishing prediction intervals rather than point estimates.
- 03
Trained asset-failure models on sparse historical labels, tuned explicitly to the cost asymmetry between a missed failure and an unnecessary inspection.
- 04
Exposed a semantic layer so regulatory and internal reporting derive from one tested definition per metric.
- Kafka
- Flink
- Iceberg
- PyTorch
- dbt
- Azure
Where it landed
Forecasts now run continuously against complete data with quantified uncertainty, and maintenance is scheduled by condition rather than calendar. Every reported figure traces back to the meter readings behind it.
A short conversation with an engineer, not a sales qualification call. If we're the wrong people for it, we'll say so and point you somewhere better.