Other Solutions

A surprising amount of operational decision-making is, mathematically, an optimisation problem dressed up as a workflow. Routing vehicles, scheduling shifts, allocating slots, packing containers, sequencing production, rostering crew, assigning cases to specialists — each of these can be expressed as an objective function, a set of constraints and a search procedure that finds a defensibly good solution within an acceptable time budget.

We use mathematical programming (LP, MIP, constraint programming), metaheuristics (simulated annealing, genetic algorithms, large-neighbourhood search) and learning-augmented optimisation depending on the structure of the problem. Where the problem is well-characterised and stable, exact methods are usually preferred. Where the problem space is vast and the objective is forgiving, heuristics scale better.

Optimisation systems live or die on their integration with the operational reality. The model that finds a globally optimal route is useless if the routes it finds violate informal constraints the planner held in their head. We surface those constraints explicitly during discovery, encode them in the model, and invest in the user interface that lets planners adjust solutions without throwing the optimisation away.

• Vehicle routing and last-mile
• Workforce scheduling and rostering
• Production sequencing and slot allocation
• Network and inventory optimisation
• Pricing and revenue optimisation
• Constraint programming for complex policies

Recommender systems have become commodity infrastructure for content and commerce, but most production deployments still leave significant value on the table. Off-the-shelf systems optimise for short-term engagement signals, ignore obvious business constraints (margin, stock, brand), and fail to handle cold-start cases gracefully. We build recommenders that combine collaborative filtering, content embeddings and contextual signals, with explicit business-objective layers that reconcile what the user wants with what the business needs.

Where appropriate we move beyond list-based recommenders into reinforcement-learning approaches that optimise long-term value rather than next-click probability, and into bandit frameworks that balance exploration and exploitation explicitly. These approaches are more demanding to operate and require disciplined offline evaluation, but for the right use cases they materially outperform their simpler cousins.

Some questions are best answered not with a forecast but with a simulation. What happens to throughput if we add a station? How does the network behave if we close a hub? What is the realistic capacity envelope of the rostered crew under different demand patterns? We build discrete-event and agent-based simulations of operational systems where direct experimentation would be prohibitively expensive or risky.

Where the simulation is good enough, it becomes a digital twin: a parallel representation of the operating system that shadows the real one in real time, supports scenario analysis, and enables optimisation against realistic dynamics rather than abstract ones. Building a twin worth that name is a substantial commitment; we are honest about what is required and equally happy to deliver smaller-scope simulation work where a full twin is not justified.

“If the question is what would happen, sometimes the right tool is a simulator, not a model.”

Many of the most useful AI products inside enterprises start as data-integration problems: the same customer, supplier, product, organisation or person represented inconsistently across systems. Entity resolution and knowledge-graph construction unify those representations, and unlock downstream products — fraud detection, customer 360, supplier risk, compliance screening — that simply cannot be built on top of fragmented data.

We build entity-resolution pipelines using a combination of deterministic rules, supervised matching, and embedding-based similarity, with human-in-the-loop adjudication for the cases the model cannot decide alone. The resulting graphs are stored in fit-for-purpose systems (Neo4j, Neptune, JanusGraph, RDF stores or relational equivalents) and exposed through query layers that downstream applications can consume.

Anomaly detection is deceptively hard. Almost any approach will produce alerts; very few will produce alerts an investigative team can actually work through. We design anomaly-detection systems with the operations team firmly in mind: prioritised queues, clear contextual evidence per alert, and explicit feedback loops so that triaged outcomes train the next generation of models.

Where the underlying problem is adversarial — fraud, abuse, security — we layer in additional defences: shadow models, drift detection on attacker behaviour, and red-team exercises that probe for the patterns our deployed models miss. The objective is a system that does not merely catch yesterday's fraud, but adapts as the fraudsters do.

Generative AI is now a credible component of internal productivity tools, customer-facing assistants and content workflows. We build deployments that are useful — grounded in your data, scoped to specific tasks, evaluated against clear quality bars — and governed: with content controls, audit trails, prompt-management discipline and the right human-in-the-loop checkpoints. We are equally happy working with hosted foundation models and with self-hosted open-weight ones, and we choose the right pattern for the sensitivity and economics of the use case.