Inside The Algorithm is a new show from Cambridge Spark, hosted by Dr Jeremy Bradley, Chief AI Officer, and each episode brings in a researcher, data scientist or technical expert working at the genuine frontier of applied AI. The conversations are substantive by design. This is not a show about AI trends or tool comparisons. It is a show about the hard problems: the mathematics, the architecture decisions, the deployment realities and the open questions that do not yet have clean answers.
In the first two episodes, Jeremy sat down with Dr Lucy Morgan, Head of Simulation at the NHS Strategy Unit, and Dr Gueorgui Mihaylov, Principal Data Scientist at Haleon. Different domains, different problems, but a surprisingly consistent underlying challenge running through both conversations.
There is a version of AI that observes, reports and recommends. And then there is a version that actually does something. The gap between those two things is where some of the most interesting and most difficult work in applied AI is happening right now, and it is where both guests are operating.
The episodes are distinct in subject matter. But the questions running through both of them are remarkably similar.
Both Lucy and Gueorgui are, on the surface, solving domain-specific problems. One is modelling kidney disease progression and dialysis capacity. The other is optimising ordering policies for a global consumer health supply chain. But when you listen closely, neither of them is really talking about the domain problem. They are talking about something harder: how do you build a system that can be trusted to act, not just to inform?
For Lucy, the challenge is reproducibility and trust. The renal replacement therapy model she built with the Midlands Kidney Network works and has been validated, and it is now being rolled out to London and the East of England, with ambitions to cover the whole of England within the next year. But every time it moves to a new region, she faces the same problem. The new organisations were not in the room when the model was built. They did not go through the workshops. They have not seen the validation process unfold in real time. And so the technical confidence that exists in the Midlands does not automatically transfer with the model.
As Lucy puts it, that trust gap is harder to bridge than any of the technical challenges.
For Gueorgui, the challenge is a different kind of gap: the distance between a working prototype and genuine operational adoption. The AI Inventory Planner that his team developed demonstrated, through repeated shadowing exercises, that it could reduce total inventory positions by 12 to 16 per cent while maintaining or improving service levels. So much so that it was shortlisted for the President's Medal of the Operational Research Society, and yet the fully automated version of the tool has not been adopted wholesale by the business.
What has happened instead is arguably more durable. The underlying mathematics, the demand forecasting logic, the stochastic scenario simulation engine, and the probabilistic approach to ordering policy have been absorbed into Haleon's global integrated business planning programme and are now running at enterprise scale. The tool did not get switched on. The thinking got embedded.
Both episodes go deep into the technical architecture, and that is where the conversations become genuinely unusual for a business podcast.
Lucy's renal model is a sequential hybrid simulation that combines two distinct methodologies chosen specifically to match the nature of the process being modelled. Kidney disease is a continuous, slow deterioration. System dynamics, with its stock-and-flow structure, is the right tool for capturing that. Kidney replacement therapy, by contrast, involves discrete patient pathways through dialysis modalities and transplantation. That calls for discrete-event simulation. The two models sit sequentially, with the output of the first feeding the input of the second.
The validation approach is similarly rigorous. Lucy used data from 2010 to 2022 to fit input distributions, then assessed how well the model's outputs represented observed system behaviour from 2022 onwards. Where the Midlands had clinicians embedded throughout the build and validation process, the national rollout has required new approaches to building that same confidence without the shared history.
Gueorgui's inventory system is an ensemble of components that most people would not expect to find inside a consumer goods company. A global demand forecasting algorithm that cross-learns between time series to produce distributional forecasts rather than point estimates. A delivery forecasting model built on survival analysis, using Weibull distributions whose parameters can be estimated by a deep neural network incorporating seasonal effects and atmospheric conditions on delivery routes.
A scenario simulation engine capable of executing hypothetical supply chain runs thousands of times, compatible with a complex set of real-world constraints around minimum order quantities, ordering frequency and expedited ordering rules.
The output of all of this is not a forecast. It is a policy: the probabilistic trigger points for normal and expedited orders that are most likely to guarantee a requested service level with minimum inventory. Moving along those ISO service level curves, the system recommends the optimal ordering strategy. And sometimes, as Gueorgui notes with evident interest, that strategy is genuinely surprising. Not wrong, just different from what experienced human planners would have done.
The most technically striking material in either episode comes when Gueorgui turns to the Golden Batch project, and it is worth dwelling on it.
Glycerol-based toothpaste is a non-Newtonian fluid. Its manufacturing process runs across approximately 30 phases, each involving different combinations of ingredient addition, mixing speed, temperature and pressure. The control variables are sensor readings coming directly from the mixers in real time. The goal is to identify the corridor of execution, across all those phases and all those parameters, that consistently produces the best quality output in the shortest time.
The fundamental difficulty is that this is a highly non-stationary process. Standard statistical process control, the kind used for stable, stationary manufacturing processes, does not apply directly. The team had to develop a dynamic version of a Shewhart control chart, adapted to account for the correlation structure between 13 or 14 control parameters evolving across phases and to handle the challenge of recognising phase transitions from noisy sensor data alone.
The result, visualised through principal component analysis, is what Jeremy aptly describes as a cylindrical hypercube: a corridor through high-dimensional space that the process must stay within to guarantee quality. When the real-time sensor readings drift outside that corridor, the system flags it. The next step, still in progress, is moving from flagging to recommending corrective action. The step beyond that is a fully closed-loop control system that acts without human intervention.
That final step is what makes this genuinely novel territory. Most AI in manufacturing advises. This project is working towards AI that controls it.
Neither Lucy nor Gueorgui is under any illusion that the hardest problems have been solved. Both, in different ways, point to the same horizon.
Lucy is actively researching how AI agents could help NHS staff interrogate models directly, running what-if scenarios without needing deep modelling expertise. The vision is a system where a planner can ask what happens to kidney replacement therapy capacity if demand rises by a given percentage over the next two years and have the simulation run and return results without a modeller in the loop. She has been exploring this with academics at the University of Exeter and frames it as fundamental to making analytical models genuinely usable across the NHS at scale.
Gueorgui's frontier question is more conceptual but no less consequential. He is interested in understanding precisely where large language model reasoning adds value in industrial AI and where it gives way to orchestration. His argument is that there is a specific threshold beyond which an LLM should not be trying to solve a problem itself but should instead be directing classical tools, solvers, forecasting engines and simulation models that are better equipped for the task. Getting that boundary wrong trades solution quality for the appearance of automation. Getting it right creates something that can genuinely reshape how complex industrial systems are managed. That distinction, he believes, will define business strategy in this space sooner than most expect.
Inventory corridors. Manufacturing control corridors. Kidney disease progression flows. Simulation models navigating from one region's trust to another. The subject matter across these two episodes could hardly be more different.
But the underlying challenge is consistent. Both guests are working on systems that need to move from producing insight to producing action and doing so in environments, healthcare and global manufacturing, where the consequences of getting it wrong are significant and the barriers to trust are correspondingly high.
That is the space Inside The Algorithm is designed to explore. If the first two episodes are any indication, the conversations ahead are going to be worth following closely.
______________________________________________
Subscribe to Data & AI Mastery so you never miss an episode of Inside The Algorithm.
If you are a data and AI leader looking to build deep technical capability across your organisation, Cambridge Spark is here to help. Explore our programmes at cambridgespark.com.