AnalyticaContact
Services
Work
Products
Company
ESC

↑↓ to move↵ to open

← All services

Service

AI & Data Science Delivery

Most of what we are asked for is a model. Most of what actually gets used is a model plus the unglamorous work around it: reconciling records that do not share an identifier, making the output explainable enough to defend, and deploying it somewhere your team can operate without us.

Two systems · no shared identifier

Record linkage between two systems with no shared identifierTwo record sets, five records each, with no key in common. Blocking reduces the 25 possible comparisons to 7 candidate pairs by requiring a coarse agreement first. Each candidate is then scored across name, date of birth and address, and pairs above the threshold are accepted. 5 matches are made. Two different people both called John Smith stay separate, while J. Smyth and Smyth John P are joined — the opposite of what comparing names alone would give.SYSTEM A · CRMSYSTEM B · BILLING EXTRACT5 records5 records0.960.910.240.880.970.190.94John Smith1984-03-02SW1AJ. Smyth1979-11-15EC2RMaria Okafor1991-07-08M1John Smith1966-01-22LS1R. Chen1988-05-30EC2RSMITH, JOHN02/03/84SW1ASmyth John P15 Nov 1979EC2RM. Okafor-Bello1991-07-08M1Chen, Rui1988-05-30EC2RJohn Smith1966-01-22LS1Two different people, both called John Smith — kept apart.“J. Smyth” and “Smyth John P” — joined. Reading the names gives you the opposite.
match acceptedcompared, rejectedthreshold 0.70Blocking is the step that makes this affordable: requiring a coarse agreement first takes 25 possible comparisons down to 7 worth making, and that ratio is what decides whether the job runs in an afternoon or not at all. The scoring is the part people expect. The blocking is the part that determines whether you can do it.

What this covers

  • Predictive and risk models on tabular, geospatial and text data
  • Record linkage and de-duplication across incompatible systems
  • Time-series forecasting with honest prediction intervals
  • Unsupervised methods where labels do not exist
  • Explainability as a deliverable, not an afterthought
  • Deployment your team can actually operate

Where this tends to start

Rarely with a clean dataset. It usually starts with several systems that each hold part of the answer and none of which agree on how an entity is identified. A meaningful share of the engagements below spent their first phase on record linkage, because until the data can be joined there is nothing to model.

What we insist on

A measurable goal agreed before modelling starts. If we cannot state what number should move and by roughly how much, the project is not ready — and saying so early is cheaper for you than finding out in month nine.

Explainability sized to the decision. A risk score nobody can interrogate does not get used, and in regulated or public-sector contexts it should not be.

What we report, and why

Accuracy is the number everyone asks for and the one that hides the most. Two models can score identically and mean completely different things by “seventy percent”.

Reliability diagram comparing an overconfident model with a calibrated onePredicted probability on the horizontal axis against observed frequency on the vertical. A perfectly calibrated model follows the diagonal. The overconfident model sits well below it: the things it calls 95 percent likely happen about 62 percent of the time. The calibrated model tracks the diagonal closely, so a 70 percent prediction from it happens roughly seven times in ten. Both models can score the same accuracy; only one produces a number you can act on. The values are illustrative.PERFECTLY CALIBRATEDsays 95%, happens 62%010PREDICTED PROBABILITY1OBSERVED FREQUENCYBoth models score the same accuracy. Only one gives you a number you can act on.
overconfidentcalibratedperfectIf you only need a ranking — who to call first — calibration barely matters. If you are going to act on the number itself, stage crews against it, order a part against it, then it is the whole game: seventy percent has to mean seven times in ten. An accuracy score cannot tell these two models apart, which is why we report this one. Values here are illustrative.

Tell us what decision you are trying to get right.

Not a discovery call about our capabilities. A conversation about the specific thing you need to predict, and whether the data you have can support it.

Start a conversation