Advertising is among the world’s largest industries, and it has quietly become one of the most automated. Programmatic advertising, where adverts are bought and sold by algorithms rather than people, exceeded $700 billion in spend in 2025 and is forecast to surpass $1 trillion by 2030.
In programmatic advertising auctions, advertisers have less than 100 milliseconds to decide how much they want to spend to show their advert, and then commit real money to their decision. At the scale of the internet, the reward for targeting well is enormous, as is the cost of getting it wrong. It is a high-speed, high-stakes environment where machine learning engineers like Peter Paton thrive.
In this respect, the most sophisticated ad-tech companies operate less like marketing businesses and more like high-frequency trading firms. Architecting these systems demands a rare combination of advanced statistics and production engineering deployed under the pressure of competitive, low-latency auctions.
Paton embodies that combination. An Imperial College London-educated mathematician, he builds models that learn from trillions of data points, distill signal from noise, and make millions of reliable decisions in real time. At Quantcast, he owns the demographics prediction models, training them on over seventeen terabytes of internet data each day to guide the targeting of tens of millions of dollars of annual advertising spend.
The Mathematical Foundation
Digital advertising is governed by a single core mechanism: every time a web page loads, an auction opens for the right to show you an advert, and closes milliseconds later. In this brief window, complex algorithms figure out a user’s interests, select the most appropriate advert and bid real money.
But the difficulty is not speed alone. In Paton’s experience, consistently calculating competitive prices when the data is vast, noisy, and constantly changing is a mathematical optimization problem before it is an engineering one.
It’s a challenge he is deeply familiar with. He completed a joint master’s in Mathematics and Computer Science at Imperial College London, graduating with First Class Honours. The program is a demanding combination that sits deliberately on the boundary between two disciplines.
Specializing in random systems, his thesis derived mathematical proofs explaining how diffusion models, the core machine learning architecture behind leading image generation models, produces unfathomable results. For this, he was awarded “Distinguished Project” status by the Department of Computing, a recognition reserved for a small number of theses outstanding in technical achievement.
Real-time Decision Making
A growing share of the modern economy runs on decisions made faster than any human could make them. The constraint they share is unforgiving: there is no human in the loop and there is no second chance to get it right.
Paton understands this well. At Quantcast, his models perform price-estimation under hard latency constraints, requiring models that are robust to uncertainty and incomplete data.
In recognition of his expertise, Paton was invited by the European Union’s Defence Innovation Scheme (EUDIS) to judge a defence hackathon in Gdańsk, Poland, last autumn. He assessed teams building drone-detection systems. “The best projects understood the real-world trade-off between accuracy and the physical constraints in fitting a model onto a small, cheap chip that can be mass-produced and deployed in the field,” he remarks.
Operating at Massive Scale
The internet generates data at a volume that breaks most conventional pipelines. Paton sees this less as an obstacle than as an opportunity: behind billions of people browsing the internet each day are learnable patterns that can inform advert targeting. The difficulty is that patterns are there, but buried in huge volumes.
The obvious answer is scale. But bigger systems, in his telling, do not automatically produce better outcomes. The events advertisers care about, such as clicks, purchases and sign-ups, can be relatively rare, which means large models can easily mistake randomness for signal.
For that reason, he places statistics and engineering on equal footing. “A strong grasp of statistics protects you from overfitting to noise. The engineering is what lets you train on enough data that the signal is actually there to be found,” Paton says. “You need both, or the system does not work.” That formulation is one of the clearest markers of his authority. He is not describing machine learning as a black box. He is describing the conditions under which it becomes reliable.
Building For Uncertainty
Taken together, the skillset demanded is holistic: the mathematical rigor to reason about uncertainty, paired with the engineering skill to build and operate real-time systems that make high-stakes decisions at scale.
What unites advertising, finance, and defence is their shape. In each, perfect decisions are impossible and the goal is the best decision the available information allows. And it is that shape, rather than any single industry, that Paton has made his domain.