WaterMark ML
An industrial water treatment context generates continuous sensor readings that operators must interpret quickly and defend later.
- Client
- WaterMark ML
- Industry
- Water treatment & industrial technology
- Discipline
- Data & Analytics
- Year
- 2025
Challenge
Readings were being reviewed manually and retrospectively. Patterns preceding process deviations were visible in hindsight but not in time to act on them.
Approach
- 01Audited the raw signal for gaps, drift and sensor faults before any modelling
- 02Established a validation split that respects time order, so results are not optimistic
- 03Compared classical baselines against learned models — the simple model wins where it can
- 04Made every prediction traceable to the inputs that drove it
Solution
A modelling pipeline producing interpretable indicators alongside predictions, so an operator sees why a reading was flagged rather than only that it was.
Technologies
- Python
- pandas
- scikit-learn
- PostgreSQL
Architecture
- Reproducible Python pipeline with versioned preprocessing
- Time-aware cross-validation and held-out evaluation
- Feature attribution surfaced with each prediction
- Exportable artefacts for integration into operational tooling
Outcome
Deviation-related patterns are surfaced as they develop, with the reasoning attached — reviewable by process engineers rather than taken on trust.
Outcomes are described qualitatively. We publish performance figures only where the client has verified and approved them.