These are direct quotes, lightly edited for length. We asked permission before publishing each one.
We gave them three years of messy sales data in a dozen spreadsheets. Six weeks later we had a demand-forecasting model that actually works. Our warehouse team now orders based on the model's weekly output, and spoilage dropped noticeably in the first month.
The document-classification tool they built for us sorts about 200 scanned letters a day into the right case folders. Before that, a paralegal spent most of Monday morning doing it by hand. The accuracy is high enough that we only need to spot-check a handful each week.
Honest bunch. They told us during the data audit that our dataset was too small for the image-recognition project we had in mind, and suggested we collect another three months of photos before starting. That saved us from wasting money on a model that would have underperformed. We came back later with better data and the result was solid.
We needed a churn-prediction model integrated with our CRM. Clarity Vision AI delivered it in five weeks, including an API endpoint our sales team queries every morning. In the first quarter, we identified 38 at-risk accounts early enough to intervene. Twenty-six of them renewed.
The weekly progress updates were a relief. Every Friday we got a short, jargon-free email explaining what had been done, what the numbers looked like and what was planned for the following week. No surprises at the end of the project.
I appreciated that they pushed back on our initial scope. We wanted real-time sentiment analysis of social media mentions, but they showed us that a twice-daily batch run would give us the same actionable insights at a fraction of the compute cost. Practical thinking.
Two projects described in detail so you can see what working with us actually looks like.
This Cardiff-based company supplies fresh produce to roughly 180 restaurants and cafés across South Wales. Their ordering was based on gut feel and last year's numbers, which meant regular over-ordering of perishable items and occasional stock-outs on popular lines.
We received three years of order history, supplier lead times and calendar data (bank holidays, school terms, local events). After cleaning and feature engineering, we trained a gradient-boosted regression model that produces weekly demand forecasts per product category.
The model went live in March 2024. Within the first quarter, the warehouse team reported measurable changes:
The model retrains automatically every Sunday night using the previous week's actuals. We monitor drift monthly and have retrained the feature set once since launch, after the client added a new product line in June.
A Swansea-area tile manufacturer was relying on two human inspectors to check every tile coming off the kiln. They caught most defects, but the pace of the line (about 14 tiles per second across two lanes) meant some slipped through. Returns from retailers were costing the company roughly £4,200 a month.
We installed two industrial cameras above the conveyor and trained a convolutional neural network on 11,000 annotated images of good and defective tiles. Defect categories included hairline cracks, glaze bubbles, edge chips and colour inconsistencies.
The system runs on a single NVIDIA Jetson Orin at the line edge. Processing latency is under 180 ms per tile, well within the line speed. When a defect is detected, a pneumatic arm diverts the tile to a reject bin and logs the defect type and timestamp.
The two inspectors were reassigned to packaging and dispatch, roles the company had been struggling to fill. Monthly returns dropped from £4,200 to about £1,150 within the first eight weeks.
Figures current as of January 2025.
We count a project as "delivered" once the model is live in the client's environment and has passed the 90-day support window without falling below the agreed accuracy threshold. The retention rate measures clients who came back for a second engagement within 18 months of the first.
We have turned down projects that we did not think would succeed, and we have recommended simpler solutions when AI was overkill. Those decisions do not show up in the numbers above, but they are a big part of why clients trust us enough to return.