AI capabilities
A list of model families says little. What matters is whether the system is built and integrated, whether it was measured against the method it replaces, and whether it still runs two years later. This page is organised around those three questions.
Sectors
Imaging and multimodal interpretation, clinical decision support, structured documentation, research workflows, dataset curation and quality auditing, and evaluation designed so a clinical reviewer will accept it.
Clinical governance is part of the team: where a system crosses from information tool into regulated medical device, and what has to be in place to survive certification, is a conversation we can have properly.
Predictive modelling and maintenance, anomaly detection, document intelligence, workflow automation, systems integration, instrumentation in hostile environments and real-time factory safety monitoring. Delivered into aluminium production, metallurgy and semiconductor manufacturing.
Crop monitoring, quality analysis, multimodal sensing and data-driven decision support, including particle and bioaerosol classification. A domain where the cost of a wrong recommendation is a season, and where explainability is therefore not decoration.
Source-grounded interpretation for museums and archives: ingestion of heterogeneous and partly analogue collections, rights-aware access, provenance and attribution, multilingual delivery, and interfaces the public uses unsupervised. The engineering is the same retrieval and document work as elsewhere; what differs is that the rights position changes record by record.
01 · Build
Nine areas of capability.
Ingestion, OCR, layout and metadata extraction across heterogeneous archives; hybrid lexical and semantic retrieval; generation that cites the passage it used. An unsupported assertion in generated output is treated as a defect.
Built and running: an eight-year corporate document archive with AI classification and analytics; an ongoing intelligent-document programme since 2023.
Classification, detection and segmentation on images that are noisy, scarce or clinically sensitive, with anatomical mapping, area quantification and observer-independent scoring.
Research reference: automated quantification of eczema lesions by anatomical region — fingertips, fingers, palms, dorsum, wrists — with the University Hospital of Basel.
Speech recognition and structured documentation from professional dictation and consultation audio; entity and relation extraction from unstructured text into schemas your systems already use.
Research reference: automated transcription and summarisation of dermatology consultations, producing structured clinical documentation with human validation — with the University Hospital of Basel.
Forecasting, anomaly detection and remaining-life estimation from sensor and process data, evaluated against the cost of a false alarm and of a missed failure.
Delivered for Onsemi, a world-leading semiconductor manufacturer: aggregation of more than fifty sensor streams into AI-driven lifetime prediction for electrical parts.
Auditing a dataset before modelling it: off-topic samples, near-duplicates, label errors and the quality problems that silently cap what any model can achieve.
Research reference: large-scale automated cleaning of large data collections — detecting off-topic samples, near-duplicates and label errors at scale.
Generation of training and test material where real data cannot be shared, with metrics for fidelity, diversity, downstream utility and privacy leakage assessed separately.
Research reference: a project on synthetic-data benchmarking, developing evaluation metrics and comparative benchmarks.
Task-scoped assistants inside a defined professional workflow: bounded tool access, explicit source attribution, a named human who approves before anything is released, and a manual fallback.
Research reference: automation of digital-forensics workflows, where each step is attributable and a qualified examiner approves the result.
Configurable 2D and 3D visualisation and immersive environments for teaching, research and technical inspection.
Instrumentation for hostile environments, real-time position and condition monitoring, and the integration that puts readings into the system of record rather than onto paper.
Delivered for the Technical University in Zvolen: an immersive 3D installation for the forestry department, built for university teaching.
02 · Evaluate
Benchmark construction is a research discipline in its own right, and it is what answers the question every client eventually asks: is this better than what we do now?
A faster workflow that lowers professional quality does not pass acceptance. Speed is never the sole criterion.
03 · Deploy
Artabel has been putting software into other organisations’ existing systems since 2004. The AI work is more recent; the integration and maintenance discipline is not.
Reusable engineering for data pipelines, APIs and connectors, role-based workflows, logging, automated testing and deployment. Built once and configured per client, so the same codebase is maintained rather than a separate copy for each.
Company-operated development, integration, testing and delivery environment with version-controlled repositories and separated test workflows. Compute and storage in Switzerland or the EU for routine AI workloads, with additional capacity added only where the workload, data location, security requirements and cost justify it.
Dataset and model version control, task-level evaluation harnesses, error analysis, human-review queues, structured logging and deployment monitoring — so drift shows up on a dashboard before a user reports it.
Mandatory export, documented interfaces, pre-approved alternative services and substitution testing. Any material change of provider, model version, data location or retention triggers reassessment.
We document hosting, maintenance and support cost and hand over a total-cost-of-ownership profile. If the maintenance cost will not fit the budget available after the project, we say so while it can still be changed.
A named owner, a deployment guide, a support and incident route and a regression suite. The measure of a good handover is that the client can change the system without us.
04 · Govern
Assurance proportionate to risk: an internal drafting aid does not need the review depth of a clinical document. But every system we ship keeps the minimum record needed to reconstruct what it did and who approved it.
How to start
Two to six weeks. One question, your data, and a written answer: what is achievable, what it would cost, and where it is likely to fail.
A working system plus the evaluation that shows whether it beats the current way of working, delivered as one package.
Someone else built it; you need a defensible verdict. Benchmark design, security, accessibility and deployment verification as a separate engagement.
Artabel joins as a named beneficiary and carries engineering work packages, with deliverables, acceptance and audit-ready records.
Next step
The interesting conversations start with the constraint that makes a project difficult — the data that cannot move, the quality bar that cannot slip, the deadline that cannot shift.