[{"data":1,"prerenderedAt":37},["ShallowReactive",2],{"content:expertise-applied-ai:en":3},{"slug":4,"number":5,"name":6,"short":7,"reveal":8,"intro":9,"problems":10,"services":15,"approach":28,"technologies":29,"notFor":36},"applied-ai","02","Applied AI","Language models and automation where they earn their keep.","Document processing, classification, assistants and decision support — evaluated against real data before anyone calls it a success.","Language models are useful for a specific set of problems: reading unstructured documents, classifying requests, drafting, searching and summarising. We integrate them into real workflows, with measurable quality, clear fallbacks and humans in the loop where it matters.",[11,12,13,14],"Staff spend hours copying data from emails, PDFs and scanned documents into other systems.","Incoming requests need to be sorted, routed or answered, and the rules are hard to write down.","Knowledge is spread across manuals, tickets and shared drives, and nobody can find anything.","An AI pilot looked impressive in a demo and never made it into production.",[16,19,22,25],{"t":17,"d":18},"Document and email processing","Extraction of structured data from invoices, orders, delivery notes and correspondence — with validation and review queues.",{"t":20,"d":21},"Classification and routing","Sorting requests, tickets or leads by intent, urgency or responsibility.",{"t":23,"d":24},"Internal search and assistants","Retrieval over your own documents with sources, permissions and honest uncertainty.",{"t":26,"d":27},"Evaluation","Test sets and metrics so you know whether a model is good enough before it touches production data.","We start with the workflow, not the model. Every AI component gets an evaluation set built from your real cases, a defined fallback when it is unsure, and logging that respects data protection. Where a simple rule does the job, we use the rule.",[30,31,32,33,34,35],"Claude & other LLM APIs","Structured outputs","Retrieval (pgvector)","Evaluation pipelines","Python","Workflow automation","We do not sell AI as a strategy. If the problem is a missing process or poor data, a model will not fix it — and we will say so.",1794663560212]