Conflux
A hazard is often five ordinary signals occurring together, not one signal crossing a threshold.
Context & problem
Single-threshold alarms alert on one sensor at a time. They miss compound hazards: situations where several weak, individually-normal signals combine into a real risk, or where a maintenance permit conflicts with equipment already running in the same zone.
Architecture
Live plant sensor data streamed in.
The interesting engineering: Five ordinary signals, one real hazard
Alerting on each specialist signal independently produces noise, since most individual signals are fine most of the time. Multiplying co-occurring weak signals instead of alerting on each in isolation is what surfaces the compound cases that threshold alarms miss.
Shift logs turned out to be an underused signal: mining code-mixed Hinglish text for equipment-bypass language surfaced risk language that structured sensor data alone would not catch. 44 tests pin the scoring behavior so a rescore change cannot silently shift what counts as risk.
Result
every plant zone, not per-sensor
weak-signal combination, not isolated thresholds
scoring logic regression coverage
What I learned
- The most valuable signal was not a new sensor. It was text nobody was reading systematically: shift-log language in a mixed register that a naive English-only pipeline would have missed entirely.
- Modeling relationships (permit-to-zone conflicts in a graph) caught a category of risk that time-series sensor data alone cannot represent.
Source & demo
No hosted demo yet. Conflux expects live plant telemetry to mean anything, so the architecture and test coverage in the repo are what to look at instead.
View on GitHubNext project
StockCraft