General questions
The basics people ask before starting a project.
Most engagements follow a two-phase pattern. Phase one is a rapid prototype: we take a representative slice of your data, build a model, and show you measurable results. This usually takes three to five weeks. Phase two covers production integration, monitoring dashboards, and staff training, which adds another two to four weeks. Smaller projects like a single classification endpoint can be done end-to-end in under three weeks. Larger programmes with multiple models and complex data pipelines have taken us up to four months, though that is the exception rather than the norm.
It depends on the task. For tabular prediction problems such as demand forecasting, we generally want at least 2,000 rows of historical data with consistent labels. Computer-vision tasks require roughly 500 labelled images per class for fine-tuning, though we can augment smaller sets with rotation, cropping, and colour-shift techniques. If you have very little data, we will tell you upfront and may suggest a rule-based approach or a pre-trained model with minimal fine-tuning rather than a full custom build.
Our largest clusters of experience are in logistics, retail, healthcare administration, and insurance. That said, the underlying techniques transfer well. We have completed projects for a fish-processing plant (defect detection on a conveyor belt), a mid-size law firm (contract clause extraction), and a chain of veterinary clinics (appointment no-show prediction). If your problem involves structured or semi-structured data and a decision that a human currently makes by pattern recognition, there is a good chance we can help.
In our experience, no. What happens more often is that repetitive, low-value tasks get automated and the people who used to do them move into higher-value work. One warehouse client reassigned three stock-counting staff to quality assurance roles after our vision system took over counting. They did not lose anyone; they just stopped doing boring work. We design systems with human-in-the-loop checkpoints specifically so your team stays in control.
Data and security
Questions about how we handle your information.
Not necessarily during the early stages. We often work with anonymised or synthetic data for the prototype phase. When real data is required for final training, we sign a data-processing agreement that specifies retention periods, access controls, and deletion procedures. All data stays within UK-based infrastructure unless you explicitly request otherwise. We hold Cyber Essentials certification and follow the ICO's guidance on AI and data protection.
By default we deploy to cloud servers in the London region (AWS eu-west-2 or Azure UK South). If your compliance requirements demand on-premise hosting, we can package models as Docker containers and deploy them on your own hardware. We have done this for two NHS-adjacent clients who could not allow patient-adjacent data to leave their network perimeter. Latency and cost implications differ between the two approaches, and we will walk you through the trade-offs before you commit.
We delete all client data from our systems within 30 days of project completion unless you ask us to retain it for ongoing model retraining. You receive a written confirmation of deletion. The trained model weights belong to you, and we hand them over in a standard format (ONNX, TensorFlow SavedModel, or PyTorch checkpoint) along with documentation on how to load and run them independently.
Pricing and process
How much it costs and how we structure engagements.
We offer two models. The first is a fixed-price prototype engagement: you pay a flat fee (typically between £4,000 and £12,000 depending on complexity) for a working proof of concept with documented accuracy metrics. If you decide to go to production, we scope a second phase with its own fixed price. The second model is a monthly retainer for ongoing model monitoring, retraining, and support, which starts at £1,200 per month. We do not bill by the hour because it creates the wrong incentives.
Before we start, we agree on a target metric: for example, "the model must classify tickets with at least 85% accuracy on a held-out test set." If we miss that target after two rounds of iteration, you owe nothing for the prototype phase. This has happened three times in our history, all in cases where the underlying data turned out to be too noisy for the task. We would rather lose a fee than deliver something that does not work.
Yes. We expose models as REST APIs with JSON input/output, which means any system that can make an HTTP request can call them. We have integrated with Salesforce, SAP, Xero, custom Django apps, and several legacy systems that only spoke SOAP. If your platform has an API or even a database we can query, integration is straightforward. We also provide webhook-based triggers for event-driven architectures.
Still have questions?
Drop us a line and we will get back to you within one working day. No sales pitch, just a straight conversation about whether AI makes sense for your situation.
Phone: +44 131 056 4068
Email: [email protected]
Address: 18 Jazlyn Brae, Old Hageneshill, Northern Ireland, MR67 7EG, United Kingdom