Software & techCode
AI engineering without the meter running
A Wales-based software company piloting a hosted open-weight service for coding assistants and agents — replacing overseas per-token APIs for day-to-day engineering, with usage logged per key for quality, capacity and cost benchmarking.
£0 per token during the pilot, with parity checks against commercial API output quality.
HealthcareData & insight
Smarter epilepsy pathways
AI-assisted triage and reporting across an EEG pathway — identifying low-value recordings before they consume specialist capacity, while clinician oversight stays central.
Target: ~30% fewer low-value recordings per 100 eligible cases, releasing 65–78 specialist staff-hours.
Life sciencesData
Faster, safer clinical AI updates
A clinical-support company shortens its model-improvement cycle — reviewing and labelling low-confidence events, updating, validating and re-accepting deployment models against accuracy, latency, privacy and safety gates.
Target: 50% faster cycles — from ~10 staff-days to no more than 5.
Social careVision & sensing
Learning from fall events, not drowning in them
Review and documentation of potential fall events in care settings — separating direct care and escalation from evidence-gathering, investigation and paperwork, with careful false-positive/negative monitoring.
Target: 50% less staff time on review and documentation, without lengthening investigations.
EnvironmentVision & sensing
Rapid response to pollution events
Water-monitoring SME combining sensor data, imagery and lab workflows — AI provides a preliminary assessment within minutes while laboratory confirmation continues in parallel.
Target: ~4 → ~2 staff-hours per event and ~30% fewer routine lab screening analyses.
EnergyData & insight
Every building gets an energy report
Automated analysis of building energy data and drafting of tailored recommendations — letting a small consultancy serve a far larger portfolio without linear hiring.
Target: 67–75% less staff time per building per month — a three- to fourfold capacity increase.
ArchitectureDocuments
Reviews and reports in half the time
Practice-structure document review and report preparation — AI-assisted checklists, summarisation and drafting, with professional sign-off unchanged.
Target: ~8 → ~4 staff-hours per representative case, at equal professional quality.
InsuranceDocuments
Advisers back with clients
Paraplanner and adviser support for advice cases — drafting, review and suitability documentation assisted by a privately hosted model that keeps client files in your control.
Target: ~8 → ~4 combined hours per typical case, with accuracy and oversight maintained.
Welsh languageDocuments
Welsh-first translation & transcription
Adapted open-weight models fine-tuned on Welsh-language corpora to assist professional translators — post-editing suggestions, terminology consistency and audio transcription.
Targets: ≥20% faster post-editing and ≥40% less transcription time, measured with chrF, COMET, HTER and WER.
EducationDocuments
More learners, same teaching hours
Differentiated learning materials and support generated around class data — growing a community school's reach without proportionally growing teacher preparation time.
Target: ~40 → ~50 learners — 25% growth without more staff hours.
AnalyticsData & insight
Never blocked by a rate limit again
High-volume data analysis SMEs running pipelines against locally hosted models — benchmarked against overseas APIs on elapsed time, cost, completion and energy, with no rate-limit stalls.
Targets: ≥70% lower cost, 50% faster processing, threefold monthly throughput.
EngineeringVision & sensing
Vision, sensors & simulation
A core capability of the platform, open to new projects: computer vision inspection, time-series and sensor analytics, simulation and engineering AI — with edge units deployed on site for live validation.
Available now for scoping with a Bangor academic lead — bring your dataset and your problem.
Research & academiaFine-Tuning
Domain-specific LLM fine-tuning on 768 GB VRAM
Training and fine-tuning 70B+ parameter open-weight models (Llama 3.3, Qwen 2.5, DeepSeek-R1 distillations) using LoRA, QLoRA, and full parameter tuning for Celtic/Welsh bilingual NLP, medical data, and scientific domains.
768 GB ECC VRAM cluster enables large model post-training and DPO alignment without cloud GPU rate limits or per-token fees.
Research & academiaAutonomous Agents
Autonomous multi-agent orchestration & discovery
Developing and evaluating collaborative multi-agent architectures (LangGraph, CrewAI, AutoGen) executing complex coding, planning, and scientific discovery tasks within a local, sandboxed environment.
Sub-millisecond local API execution enables high-throughput agent feedback loops, tool execution, and benchmark evaluation.
Research & academiaEdge AI & Robotics
Embodied edge AI & disconnected field trials
Deploying quantized vision-language models and reasoning agents to our four portable NVIDIA DGX Spark units for autonomous marine, environmental, and robotic field experiments operating without internet.
128 GB unified memory per portable unit allows 32B vision models to run in real-time on-site in remote Welsh environments.