What matters
The problem comes before the model.
- Features that were genuinely available at prediction time, without future leakage.
- Alignment with delayed and noisy laboratory assays.
- OOD detection and the ability to abstain instead of always producing a number.
Lab Intelligence approach
Sources, assumptions, checks, versions, metrics and limitations remain visible. A more complex technique is used only when it adds value over an understandable baseline.
How the method works →Related workflows
Examples already documented.
These pages describe implementations or experiments from the laboratory; they do not imply that the same result transfers automatically to another context.
Virtual Analyzer
Experimental soft sensor on synthetic data (simulated data): estimates product quality between laboratory assays with delay, noise, point-in-time controls, OOD detection and measurable abstention.
Open project →WF-14Process Sentinel
Detects anomalies and faults on a generic dynamic chemical process using a frozen 28-day synthetic data (simulated data) benchmark, isolated truth and a drift-aware PCA + process-residual ensemble.
Open project →Want to discuss a similar case?
Describe the problem and expected result in non-confidential terms. The first step is to identify the data, criteria and limitations that would make the case verifiable.