Injuries and fatalities in UK construction and utilities have plateaued for eight years. In 2021 alone, 15 fieldworkers died and more than 2,000 were injured so badly that they couldn’t return to work. The strategies in use aren’t working.

At FYLD I led a cross-discipline team - alongside 7 industry partners - to test whether a machine learning model could accurately predict and amplify worksite safety. The plan: combine ML and augmented reality so every fieldworker makes the safest choice, every day.

Partners: Scotia Gas Networks (SGN), National Grid, Thames Water, Kier, Morrison Water Services, Morrison Energy Services, Yorkshire Water, Cadent

The problem

The Accident Triangle — sometimes drawn as an iceberg — says that for every visible injury or fatality there are many more unreported incidents and near misses underneath. Capture the indicators and you can run experiments to reduce the events that lead to harm.

The trouble is, those indicators are massively under-reported. In 2021:

• SGN captured an estimated 2% of the indicator events they should have seen — a shortfall of around 85,000 reports
• Thames Water — five years further along on the safety journey — captured ~60%, still short by ~80,000
• Across the utilities industry, the estimated shortfall ran to 33.6 million indicator events

Without that data, no model — statistical or otherwise — has anything reliable to learn from.

What each role needed

Approach

FYLD was awarded £60k and twelve weeks to assess feasibility. I ran discovery end-to-end and partnered with a senior ML engineer on a proof-of-concept model.

I interviewed HSE leaders, operational directors and fieldworkers across Kier, Thames Water, Yorkshire Water, National Grid, Morrison Group, SGN, Cadent and several utility subcontractors. Every company followed broadly similar safety procedures — toolbox talks, the authority to shut down operations, safety reporting baked into operational KPIs — but the data each company produced was patchy, single-perspective, and impossible to compare.

On-site with Cadent — one of seven utility partners in the discovery phase

What I designed

01

Cross-industry discovery study

Interviews across seven utility operators and their subcontractors. Mapped the common procedures, the inconsistencies, and the moments where reporting friction lost the data we needed to predict anything.

For: HSE managers · Utility companies

02

Safety matrix

Co-designed with HSE managers, the matrix scored incident outcomes consistently across companies for the first time — turning subjective recall into a structured, comparable signal the model could learn from.

For: HSE managers · ML model

03

Data preparation for ML training

Manually parsed individual incident reports, employment records and open-source data. Identified the key factors and outcomes the model could learn from. Prepared two datasets — one for training, one to test capability.

For: ML engineering

Every record traces back to a fieldworker in PPE with a phone — the source of every data point we trained on

04

Stochastic ML model (with ML eng)

Tested 15 candidate models with a senior ML engineer. Two passed the 50% recall threshold on multiple runs — enough signal to justify Alpha investment.

For: FYLD R&D · OFGEM

05

Patent-pending safety reporting form

Redesigned the FYLD reporting form to capture multi-perspective accounts, root cause and impact — closing the data quality gap discovered in feasibility. FYLD named me as inventor on the patent application.

For: Fieldworkers · HSE managers · ML model

06

OFGEM SIF Alpha application

Authored a 6,000-word application with a costed project plan addressing every challenge surfaced in feasibility — automated data capture, real-time intervention via AR, and a shared safety reporting protocol across operators.

For: OFGEM · FYLD

Vision: real-time AR overlays surfacing risk where work happens

Outcomes

What changed for each role