Multi-modal correlation and anomaly detection on satellite data, at facility scale, over years
Given repeated observation of a fixed location by sensors that measure different physical properties on irregular revisit schedules, establish what normal looks like there, then decide whether current activity departs from it, by how much, with a confidence you are willing to defend.
Relating SAR, optical, RF, and thermal observations of the same place, when each sees a different property and none is ground truth for the others.
Baselines built from years of a facility's own history, where the interesting signal is a deviation from a pattern rather than the presence of an object.
Producing output that carries stated confidence and traceable reasoning, because a score with no explanation cannot enter an assessment. Machine learning and classical methods coexist here by design, not as a transitional stage.
Five things that make this a real problem
Revisit is driven by orbits, tasking, and weather, so the series has irregular gaps.
There is no annotated corpus of pre-deployment activity to train against, which puts this closer to unsupervised deviation detection than to classification.
Facilities are refurbished, seasons change, routines are rescheduled, and separating a meaningful departure from an ordinary one is most of the difficulty.
A facility that has gone quieter than its history predicts can matter as much as one that is busier, and a model built only to detect presence will miss it.
Our output is read by an analyst who will be asked to justify a conclusion, so a score that cannot be interrogated is not usable, whatever its accuracy.
The application is a ministry-level strategic assessment in the Indicators and Warnings function upstream of targeting. Our users produce judgements about force posture and intent, and the platform informs that judgement rather than making it.
We are early-stage: pre-product, pre-revenue, small team. The work at this stage is architectural, and decisions taken now will form the foundation of the product for years to come.
People who work close to the data and treat a result as provisional until they understand why it holds.
People who cannot leave a problem half-solved, and who will stay with a result that almost works until they understand why it does not.
Depth in at least one of geospatial processing, multi-modal machine learning, or time series anomaly detection, with enough interest in the others to argue about them, and a habit of stating confidence honestly, including when it is low.
Comfort with a serious domain approached without theatre. Nobody here talks about the work in tactical language.
No roles are posted. Write to us anyway
We would rather read a specific note than publish a generic job description.
Send a CV, a profile link, or both, with a short note on what you want to work on, and optionally one thing you have built close to data and what was difficult about it.
Submit a general application