Case study, transport, synthetic data

From raw sensor feed to a board level view of cost and risk.

An end to end analytics platform for the Scottish strategic road network, built to show the full job: messy data in, tested and modelled, and a clear answer out. Every step is open to inspect.

Explore the live platform Read the method

Project facts

Scope: 74 sites, 9 roads
Period: 2023 to 2025
Records: 100,000 hourly
Data: Synthetic traffic

95.70%Records passing every quality rule
63.7%Collisions caught in the riskiest 5% of hours
£1,353.3mCentral annual cost estimate

The challenge

Raw sensor data is never ready to use.

Road sensors produce a reading every hour, at every site. Some readings are missing, some are impossible, and some sites drift out of calibration. Report straight from that feed and the totals are wrong before anyone looks at them.

The questions that matter to decision makers are not in the raw data at all: which roads cost the most in delays, collisions, and carbon, and when is the next collision most likely. Answering them means cleaning, modelling, predicting, and costing, in that order, with every step checked.

That is the same problem most businesses have with their spreadsheets, at a larger scale.

What was built

Six stages, each one tested before the next.

GenerateSeeded, reproducible generation of hourly sensor readings, so every run gives the same result.Python
Transform33 models take raw readings from bronze to business-ready marts.dbt and DuckDB
Check quality120 automated tests and 12 rules, each with a clear disposition.dbt tests
StoreA star schema: 8 dimensions, 3 fact tables, and 5 aggregates.PostgreSQL
PredictGradient boosting forecasts collision risk from advance conditions.scikit-learn
Report232 measures turn it all into a board level view of cost and risk.Power BI

What it shows

Three numbers, each with its working shown.

95.70%
Records passing every quality rule
96,652 of 100,000. The rest are quarantined, imputed, or flagged, never silently deleted.
63.7%
Collisions caught in the riskiest 5% of hours
From weather, demand, and disruption known in advance. An upper bound, because the data is synthetic.
£1,353.3m
Central annual cost estimate
Range of roughly £1.05bn to £1.65bn, using DfT TAG casualty values and DESNZ carbon values.

Inside the platform

The environment and net zero page.

Annual CO₂, carbon cost, emissions intensity and roadside air quality, filtered by corridor, city, year and urbanisation. Built on synthetic data.

Environment and net zero dashboard: annual CO2, carbon cost, CO2 intensity, mean air quality index and freight CO2 share, with monthly CO2 by vehicle class, emissions by corridor and an air quality heatmap by hour and month — Environment and net zero: emissions, carbon cost and air quality by hour and month

Explore the live platform

Stated limits

What this platform cannot tell you.

Good analysis says where it stops. These limits are published on the platform itself, and your project gets the same honesty.

Synthetic data proves the method works, not what happens on real roads.
Carbon is the largest cost and carries a plus or minus 50% uncertainty.
There is no origin to destination routing.
Space mean speeds are used rather than arithmetic means.
The risk model only uses conditions known in advance.
Road segments are treated independently.
The site is a static build of the final pipeline run.

For your business

The same discipline, applied to your data.

Quality rules with dispositions: Your spreadsheets get checked and fixed, and nothing disappears without a record.
A tested, layered data model: Your reports refresh reliably, and the numbers match across every dashboard.
Measures defined once: One agreed definition for revenue, margin, or turnaround time, used everywhere.
Honest limits, stated upfront: You know exactly what your data can and cannot tell you, before you rely on it.

Built on generated data, not client data. Road names, cities, and appraisal values are real and publicly sourced; every traffic measurement is synthetic. Not affiliated with Transport Scotland.

Next step

Want this level of care on your own data?

Start with the free Health Check, or book a 30 minute review and we will look at what your data can tell you.

Book a free review Check your data free

Raw sensor data is never ready to use.

That is the same problem most businesses have with their spreadsheets, at a larger scale.

Six stages, each one tested before the next.

GenerateSeeded, reproducible generation of hourly sensor readings, so every run gives the same result.Python

Transform33 models take raw readings from bronze to business-ready marts.dbt and DuckDB

Check quality120 automated tests and 12 rules, each with a clear disposition.dbt tests

StoreA star schema: 8 dimensions, 3 fact tables, and 5 aggregates.PostgreSQL

PredictGradient boosting forecasts collision risk from advance conditions.scikit-learn

Report232 measures turn it all into a board level view of cost and risk.Power BI

Three numbers, each with its working shown.

95.70%

Records passing every quality rule

96,652 of 100,000. The rest are quarantined, imputed, or flagged, never silently deleted.

63.7%

Collisions caught in the riskiest 5% of hours

From weather, demand, and disruption known in advance. An upper bound, because the data is synthetic.

£1,353.3m

Central annual cost estimate

Range of roughly £1.05bn to £1.65bn, using DfT TAG casualty values and DESNZ carbon values.

What this platform cannot tell you.

Good analysis says where it stops. These limits are published on the platform itself, and your project gets the same honesty.

Synthetic data proves the method works, not what happens on real roads.

Carbon is the largest cost and carries a plus or minus 50% uncertainty.

There is no origin to destination routing.

Space mean speeds are used rather than arithmetic means.

The risk model only uses conditions known in advance.

Road segments are treated independently.

The site is a static build of the final pipeline run.

The same discipline, applied to your data.

Quality rules with dispositions

Your spreadsheets get checked and fixed, and nothing disappears without a record.

A tested, layered data model

Your reports refresh reliably, and the numbers match across every dashboard.

Measures defined once

One agreed definition for revenue, margin, or turnaround time, used everywhere.

Honest limits, stated upfront

You know exactly what your data can and cannot tell you, before you rely on it.

Built on generated data, not client data. Road names, cities, and appraisal values are real and publicly sourced; every traffic measurement is synthetic. Not affiliated with Transport Scotland.