Until recently, around 10% of the UK’s water network relied on members of the public to report illegal sewage spills. At the NWG Innovation Festival 2022, I led “The Spillage Idiots” to win Best Pitch by showing how machine learning could close that gap — and dramatically cut the time it takes to verify a spill once it’s reported.

The problem

Yorkshire Water’s discovery brief surfaced a single hard number: each spill analyst spent an average of 36 hours validating one spill. With 478 of 1,067 sites still requiring some manual validation that a spill was even occurring, the verification queue was the bottleneck — and every hour in that queue raised the risk of regulatory fines and harm to people, property and the environment.

Two groups had to win for the strategy to work.

What each role needed

Approach

I worked with a senior ML engineer overnight on data from 408 sites across the Yorkshire Water network. Before any modelling we had to organise the inputs by how trustworthy each signal was — only then could a stochastic model weigh them sensibly.

Persona-needs mapping — water company, regulator, public — driving the model brief

Outcomes, deliverables, and the seven-metric verification matrix

What I designed

01

Three-tier data confidence framework

Sorted every available signal into three buckets — absolute (manual checks, maintenance logs), high (on-site telemetry, nearby rainfall, spill duration data) and low (nearby sites, historic weather trends, nearby flooding). Gave the model a structured way to weigh the inputs and a structured way to explain its predictions.

For: ML engineer · Yorkshire Water analysts

02

Stochastic ML proof of concept

Trained on data from 50% of the sites, tested on the other half. The model reached ~90% confidence and >80% accuracy that a spill had occurred — enough to demonstrate viability, with a clear path to higher accuracy through more data and longer training.

For: Yorkshire Water

03

Phased rollout plan

A practical adoption path Yorkshire Water could pursue — model in the loop alongside analysts, then progressive automation as accuracy improved. Made the proof of concept land as a roadmap, not a science project.

For: Yorkshire Water leadership

04

The pitch

Authored and delivered the final pitch to a room of water-industry experts. Framed the work around regulatory risk and public harm — not the technology — so the audience could see the business case before the model.

For: Hackathon judges · Yorkshire Water · Wider water industry

Pitching to the room — framed around regulatory risk and public harm, not the model

Outcomes

What changed for each role