Service details
Predictive analytics
We take your historical records, whether they sit in spreadsheets, a SQL database or a warehouse like BigQuery, and build models that forecast future outcomes. Common use cases we have delivered include:
- Weekly demand forecasting for a Cardiff-based food distributor, reducing spoilage by 18 % in the first quarter.
- Customer churn scoring for a SaaS company with 12,000 subscribers, identifying at-risk accounts 30 days before cancellation.
- Predictive maintenance alerts for a fleet of 140 delivery vans, cutting unplanned downtime by roughly a third.
A typical project runs four to eight weeks. We start by profiling and cleaning the data, then test several model architectures, select the best performer and deploy it behind a REST API your existing systems can call. You receive a live dashboard and a written model card explaining accuracy, limitations and retraining schedule.
Computer vision
We design systems that interpret images and video streams. Our team handles the full pipeline: camera selection advice, image annotation, model training, edge deployment and ongoing monitoring.
Recent projects include a defect-detection system for a ceramics manufacturer in Swansea. The system inspects tiles on a conveyor belt at 12 tiles per second and flags cracks, chips or glaze inconsistencies with 97.3 % recall. False positives run at about 1.8 %, which the line operators confirmed was well within tolerance.
We also built a document-classification tool for a legal firm that sorts scanned correspondence into 14 categories, saving their admin team roughly six hours a week. The model runs on a single GPU workstation in their server room; no cloud dependency, no recurring compute bill.
Natural language processing
Text is messy. Spelling varies, abbreviations multiply, and context matters enormously. We fine-tune language models on your specific vocabulary so they understand what your customers and colleagues actually write.
One project we are particularly proud of involved a Welsh council's housing department. Residents submit repair requests by email, and the old process required a clerk to read each one, categorise it and assign it to the right trade. Our classifier now handles that triage automatically for about 85 % of incoming emails, routing plumbing issues to plumbers, electrical faults to electricians, and so on. The remaining 15 % go to a human reviewer. Average processing time dropped from four hours to eleven minutes.
We also offer sentiment analysis, entity extraction and summarisation services for companies that need to process large volumes of survey responses or social-media mentions.
Custom model development
Sometimes the problem does not fit neatly into one of the categories above. A logistics company needed a route-optimisation engine that combined demand forecasts with real-time traffic data. A healthcare startup wanted an anomaly-detection model for wearable-sensor readings. These projects required bespoke architectures and custom training loops.
We scope custom work carefully. You get a written specification before we write a single line of training code, including expected accuracy ranges, data requirements and a delivery timeline broken into fortnightly milestones. If we cannot meet a milestone, you hear about it that same week, not at the final demo.
How a project moves from idea to production
Five phases, each with a clear deliverable you can review before the next begins.
Data audit
We spend two days examining your data sources, formats and quality. You receive a written report with a go/no-go recommendation. This phase is free.
Specification
We define success metrics, agree on data-handling rules and write a detailed project plan. Both sides sign off before development starts.
Development
We build, train and validate the model. You see progress every two weeks in a short video call with live demos.
Deployment
The model goes live in your environment: cloud, on-premise server or edge device. We handle containerisation, CI/CD and monitoring setup.
Support
Ninety days of included monitoring and retraining. After that, optional maintenance contracts start at a flat monthly fee.
Pricing approach
Fixed fees, not timesheets
We quote a fixed price for each phase of work. If the scope stays the same, the price stays the same. If you want to add features mid-project, we issue a change-order quote within two business days and wait for your approval before proceeding.
Typical project costs range from £8,000 for a focused analytics model to £45,000 for a full computer-vision pipeline with edge deployment. The data audit is always free.
We do not offer hourly consulting. Experience has taught us that hourly billing creates misaligned incentives: we would earn more by working slowly, and that is not how we want to operate.
Request a quote