Work
Engagements where being wrong was expensive.
Each of these is written the way we would explain it to another engineer: what the problem actually was, what we built, what the data would and would not support, and what changed as a result.
01
Federal government · Homeland security · Critical infrastructure
2018 – present
Predicting how a hurricane cascades through US critical infrastructure
Record linkage at national scale, then a forecast of how a storm moves through what it links — delivered as a platform, as a subcontractor, since 2018.
Cybersecurity and Infrastructure Security Agency (CISA), US Department of Homeland Security
2018 –
continuing, as a subcontractor
02
Energy · Utilities
2 years
Predicting where a storm will break the grid — five days out
A two-stage model that forecasts damage by district, then converts that into the crews and materials to pre-stage.
A major US investor-owned electric utility
15%
reduction in outage times
03
Energy · Environment · Public sector
38 years of records, 1975–2013
Four decades of pipeline incidents, and the gap nobody had measured
Roughly 62,000 incidents across 38 years. 94% were reported the same day. Only 20% were resolved the same day — a gap that had never been quantified.
Government of Alberta
~62,000
incidents analysed
04
Commercial real estate · Hospitality
2016, then multi-year
Matching hotels to buyers, in a market where almost nobody rates anything
A recommender built on a ratings matrix that was 99.35% empty — and the honest bake-off that picked the algorithm, including the one that lost.
Jones Lang LaSalle (JLL)
3 months → instant
to produce a buyer shortlist
05
Automotive · Telematics
19,000 km over 11 days
Telling drivers apart from the way they drive
Sensor telemetry from an 11-day, 19,000-kilometre expedition, reduced to a behavioural signature that identifies who is behind the wheel.
A global automotive manufacturer
78%
identification accuracy
06
Mining · Industrial
Not disclosed
Predicting equipment failure before it stops the site
Remaining-useful-life models over historical part-failure data, plus unsupervised stress-level estimation on components with no failure history.
A mining and heavy-equipment operator
−23%
unscheduled downtime
What are you trying to get right?
Tell us the decision you keep having to make with less certainty than you would like. We will tell you honestly whether the data you already have can support it — and if it cannot, we will say so on the first call rather than the third.
Start a conversation