Predictive modelling
Time-series forecasting, demand prediction, churn scoring. We train on your historical data and validate against hold-out sets you choose, so the accuracy numbers mean something before you commit to production.
0
Models deployed
0%
Average accuracy
0+
Compute hours saved per client
0
Industries served
A small team that ships fast
We are six engineers based in Wales. Three of us came from academic research labs; the other three spent years building production systems at scale. That mix matters: we know the theory, and we know what breaks at 2 a.m. when real traffic hits your API.
Every project gets a named lead engineer who stays with it from scoping through deployment. You will never explain your data pipeline twice to a new face. Our median time from signed brief to production-ready model is nine weeks. Complex computer-vision or NLP pipelines take longer, typically twelve to sixteen weeks, and we will tell you that up front rather than pretending otherwise.
What we build
Each engagement starts with your data, not our product catalogue. Here are the areas where we have deep experience.
Computer vision
Defect detection on production lines, document parsing from scanned images, satellite imagery classification. We handle annotation, training, and edge deployment to NVIDIA Jetson or equivalent hardware.
Natural language processing
Sentiment analysis, entity extraction, document summarisation, and retrieval-augmented generation pipelines. We fine-tune open-weight language models on your domain corpus so answers stay grounded in your knowledge base.
Data pipeline engineering
A model is only as reliable as its input. We design ingestion, cleaning, and feature-engineering pipelines in Apache Airflow or Prefect, with automated drift monitoring that pages your team when distributions shift.
AI safety and audit
Bias testing, explainability reports for regulators, GDPR-compliant model documentation. If you already have a model in production and need a third-party review, we can run a two-week audit and deliver a written risk assessment.
MLOps and maintenance
Continuous training loops, A/B deployment with canary rollouts, model versioning in MLflow. We can manage ongoing retraining on a monthly retainer or train your internal team to take over within 90 days.
How a typical engagement works
Four phases, each with a concrete deliverable you can evaluate before we move forward.
1. Data discovery (week 1)
We connect to your data sources, profile schema quality, and identify which features are likely to carry signal. You receive a written feasibility memo with expected accuracy ranges and a list of data gaps we need filled.
2. Prototype (weeks 2–4)
A working model in a Jupyter notebook, tested against your validation criteria. We walk your team through every decision: why this architecture, why these hyperparameters, what the failure modes look like.
3. Production build (weeks 5–8)
The notebook becomes a containerised API with logging, health checks, and automated tests. We deploy to your cloud account so you own the infrastructure from day one.
4. Handover and monitoring (week 9+)
Documentation, a runbook for common failure scenarios, and a Grafana dashboard tracking prediction latency, throughput, and data drift. Optional ongoing retainer for retraining and support.
Questions we hear often
How much data do we need before a project is viable?
It depends heavily on the task. A tabular churn model can work well with a few thousand labelled rows. A custom image classifier usually needs at least 500 annotated images per class, though data augmentation can stretch smaller sets. During the discovery phase we give you an honest assessment; if your data is too thin, we will say so rather than start a project destined to underperform.
Do you work with companies outside Wales?
Yes. About half our clients are elsewhere in the UK, and we have delivered projects for teams in Ireland and the Netherlands. Most collaboration happens over video calls and shared Git repositories, so location rarely matters. For on-site workshops we charge travel at cost.
What does a project typically cost?
A focused predictive model with clean data starts around £12,000. Larger NLP or computer-vision systems with custom annotation and edge deployment tend to fall between £30,000 and £65,000. We quote fixed price per phase so you can stop after any deliverable without further obligation.
Who owns the trained model and code?
You do. All source code, model weights, and documentation are delivered to your repository under a full-transfer IP clause. We retain no copies after project close unless you ask us to provide ongoing support.
Can you integrate with our existing systems?
We have integrated models into SAP, Salesforce, custom Django and Rails applications, and various ERP platforms via REST APIs and message queues. If your stack uses something unusual, we will confirm compatibility during the discovery phase before any commitment.
Let us know what you are working on
Describe the problem in a few sentences. We will respond within one working day with initial thoughts and, if it looks like a fit, suggest a 30-minute call to dig into the details.
Address: 33 Richard Grange, McKenzie-over-Ullrich, Wales, OH6 8TP, United Kingdom
Phone: 01172 20909
Email: [email protected]