We build models that classify, forecast and automate real processes for mid-size companies. No hype decks. You send us a problem and a dataset; we return a working prototype in under three weeks.
Send us your dataset challenge
Each card summarises a real engagement. We anonymise client names unless they opt in.
We replaced their spreadsheet-based ordering with a gradient-boosted model trained on 14 months of POS data. Waste dropped 23 % in the first quarter. The model retrains weekly on fresh sales figures.
time-series · XGBoost
A convolutional network running on an edge device flags solder defects at 12 frames per second. False-positive rate sits below 0.8 %, which was the threshold the client needed to retire manual inspection on two lines.
computer vision · edge inference
Fine-tuned a transformer language model on 60,000 historic support tickets. It now routes incoming messages to the correct team with 91 % accuracy and tags urgency so that frustrated customers get a response within 15 minutes.
NLP · fine-tuned LLMFive stages, typically completed in 12 to 20 working days.
We spend 45 minutes understanding what you want the model to do, what data you have, and what "good enough" looks like in your context. No charge for this call.
You share a sample. We profile it: row counts, missing values, class balance, feature distributions. If the data can't support the goal, we say so here rather than later.
We train candidate models, benchmark them against a hold-out set, and document the trade-offs. You receive a short report comparing two or three approaches with accuracy, latency and cost estimates.
The chosen model gets wrapped in an API or embedded in your existing pipeline. We write monitoring hooks so you can see prediction drift before it becomes a problem.
Full documentation, a recorded walkthrough, and access to the training code. If you want ongoing retraining or model updates, we offer a monthly retainer starting at £800.
Four capability areas. Most projects combine two of them.
Revenue forecasting, churn prediction, equipment failure probability. We work with structured tabular data and typically use gradient-boosted trees or Bayesian regression, depending on how much historical data is available. Minimum viable dataset: around 5,000 labelled rows.
Document classification, entity extraction, summarisation, and sentiment analysis. We fine-tune open-weight language models so your data never leaves your infrastructure. Typical training time on a single GPU: four to eight hours.
Defect detection, object counting, document OCR. We deploy models on edge hardware (Jetson, Coral) when latency matters, or on cloud GPU instances for batch processing. Annotation support is included in the project fee.
A fourth area, recommendation systems, is something we take on selectively. If your catalogue has fewer than 500 items, collaborative filtering rarely outperforms a hand-tuned rule set, and we will tell you that upfront.
If yours isn't here, the contact form is just below.
Describe your problem. We respond within one working day.
