A selection of production AI systems we've designed, shipped and continue to operate. Names withheld where contracts ask, numbers verified where they don't.
Replaced manual QA on an automotive line with an edge neural pipeline running on commodity GPUs. Throughput up 3×, scrap rate down 41%.
1,800 SKUs across 220 stores. A hierarchical reconciled model with promotional uplift modelling, deployed in 14 weeks.
Real-time monitoring of brand voice across social, review and support channels. Crisis early-warning routed straight into PR ops.
Curated ontology of 180k clinical concepts powering an explainable triage assistant. Reduced average intake time by 38%.
Every figure on this page is one we would defend in a steering committee. That sounds obvious; it is not. The standard industry case study reports the headline lift against a baseline carefully chosen to make the lift look maximal — last year's worst month, the weakest rules-based rival, a metric only the model improves. We do not write those. Each metric here was agreed with the client's analytics team before the project began, recomputed on holdout data the model never saw, and tracked through go-live and the first full quarter of operations. Where a number disappointed, we either fixed the model or the case study did not get written.
That stance has consequences. It means projects sometimes ship without a publishable case study because the lift, while real, was not portfolio-worthy. It means our case studies are a fraction of what we have delivered, not a comprehensive catalogue. And it means the studies that do appear are unusually conservative — many clients have asked us to upgrade the headline figure to the actual production number once the dust settled. We decline. The number that mattered was the one estimated before deployment; the larger number that emerged afterwards is a bonus, not the claim.
Yes, after NDA. Most of our published case studies are anonymised at the client's request; we can introduce named references once a mutual NDA is in place.
Only metrics agreed with the client's analytics team before the project began, recomputed on holdout data the model never saw and tracked through go-live and the first full quarter of operations.
The vision QA programme reached production in approximately 16 weeks. The hierarchical demand model deployed in 14 weeks across 1,800 SKUs and 220 stores. The sentiment radar across 14 languages went live in 10 weeks. The clinical knowledge graph reached triage pilot in 22 weeks.
Yes — these are illustrative. We also deliver in financial services, logistics, telecoms and the public sector. The Industries page lists every vertical we support.
Mid-sized to large enterprise (typically 500+ employees) and venture-backed deep-tech. We are not a fit for pre-product startups looking for a generic AI assistant.
A case study is only useful if its numbers can be trusted. Every figure on this page follows the same four-step evidence protocol, agreed with the client's own analytics team before a single model is trained.
We record the current process performance — defect escape rate, forecast error, analyst hours — over a representative period, so improvement is measured against reality, not against a strawman.
Accuracy is reported on a time-separated holdout set that is locked before training starts. No tuning, feature selection or threshold setting ever touches it.
Models run silently alongside the existing process for weeks. We compare decisions side by side and only switch over when the client's owners sign off.
Published numbers are re-confirmed after a full quarter in production, because a model that degrades after launch is not a success story.
Bring the messy reality. We'll prototype a model in weeks, not quarters.
