Getting started
It depends on the task. For classification problems like sorting support tickets into categories, a few hundred labelled examples are often enough to build a useful first model. Time-series forecasting generally needs at least two full seasonal cycles of historical data, so if your business has strong annual patterns, two years of records is a good starting point.
During the discovery sprint we assess exactly what you have and tell you whether it is sufficient. If it is not, we can help you set up collection processes so you are ready in a few months rather than a few years.
Most real-world data is messy. Missing fields, inconsistent formats, duplicate records: we see these in nearly every project. Part of our build phase is a data-cleaning pipeline that standardises inputs automatically. We also flag systemic quality issues and recommend fixes at the source, so the data improves over time rather than getting cleaned up after the fact every month.
Not necessarily. We handle all the engineering. What we do need is someone on your team who understands the business process well enough to answer questions like "what counts as a late delivery?" or "which customers are high-priority?" That person does not need to write code; they need domain knowledge and the authority to make decisions about how the system should behave.
The discovery sprint takes two weeks. A full build-and-deploy project usually runs four to ten weeks after that, depending on scope and data readiness. Dashboard projects tend to be on the shorter end; predictive models that need careful validation against business outcomes tend to be longer. We give you a specific timeline estimate at the end of the discovery sprint, once we understand your data and requirements.
Technical questions
Our core stack is Python for data processing and model training, with scikit-learn, XGBoost, and PyTorch as the main libraries depending on the problem. For APIs we use FastAPI or Flask. Dashboards are built in Metabase, Grafana, or custom React apps. Infrastructure runs on AWS or Azure, whichever your organisation already uses. We avoid introducing new cloud providers unless there is a strong reason.
That is a core design goal. We deploy models as REST API endpoints, which means any system that can make an HTTP request can call them. We have integrated with Salesforce, SAP, Xero, HubSpot, and various in-house platforms. If your system exports CSV files on a schedule, we can work with that too. During discovery we map out the integration points and test connectivity before writing any model code.
All data is encrypted in transit (TLS 1.2 or higher) and at rest (AES-256). We work within your cloud environment whenever possible, so data never leaves your control. When we need to pull a sample for local development, we anonymise it first. Our team members sign individual NDAs, and we can work under your organisation's data processing agreement if you have one.
Model drift is normal. Customer behaviour changes, product lines shift, economic conditions fluctuate. We set up automated monitoring that compares the model's predictions against actual outcomes on a rolling basis. When accuracy drops below a threshold you define, the system alerts us (or your team, if you prefer to manage retraining internally). Retraining typically takes a few hours and can be scheduled quarterly or triggered on demand.
Commercial questions
Yes. All custom code, trained model weights, and documentation we produce during a project engagement belong to you. We retain the right to reuse general-purpose utility functions and open-source libraries (which you also have access to), but the specific models trained on your data and the application logic are yours. You can take them to another provider or maintain them in-house if you choose.
The discovery sprint is £3,500, fixed price, and takes two weeks. If the prototype we build during that sprint does not demonstrate measurable value, as judged by you, we waive the fee entirely. About 85% of discovery sprints convert into full projects, and the remaining 15% usually end because the data is not ready yet rather than because the idea lacks merit. In those cases we provide a written plan for getting the data into shape.
Yes, with 30 days written notice. There is no minimum commitment beyond the first month. We find that most retained clients stay for at least six months because the ongoing monitoring and incremental improvements compound in value, but we do not lock you into a long contract.
We do, though the majority of our clients are UK-based. For international engagements we work remotely over video calls and shared project boards. Time-zone overlap is important for the discovery phase, so we are most comfortable with clients in Europe, the Middle East, and Africa. If you are further afield, we can make it work with some schedule flexibility on both sides.
Still have questions?
Drop us a message and we will get back to you within one working day. No sales pitch, just a straight conversation about whether AI software makes sense for your situation.
Get in touch