Advancing AI research: fine-tuning, autonomous agents & joint discovery
Bangor University opens its enterprise Blackwell compute facility to academic researchers, doctoral scholars, and collaborative research programmes across Wales and beyond.
High-performance compute without national queue bottlenecks
UK academic researchers often face lengthy queuing times on national Tier-1/Tier-2 supercomputing clusters or unsustainable cloud costs when training and testing emerging models.
The North Wales AI Platform delivers a modern, agile alternative: an enterprise 8-GPU Blackwell cluster engineered specifically for modern transformer architectures, continuous inference, parameter-efficient fine-tuning (PEFT), and multi-agent loops.
Led by Professor Zengbo Wang at Bangor University's School of Computer Science and Engineering, our doors are open to university academics, research fellows, and PhD researchers seeking collaborative compute, co-investigator partnerships, and joint grant bidding across Wales and beyond.
Four pillars of frontier AI exploration
From fine-tuning open foundation models to developing self-reflective autonomous agents and field-deployable edge intelligence.
LLM Fine-Tuning & Domain Adaptation
Full parameter fine-tuning for 7B–14B models and LoRA/QLoRA parameter-efficient adaptation for 32B–70B+ foundation models (including Llama 3.3, Qwen 2.5, Mistral, and DeepSeek-R1 distillations).
- Bilingual Welsh & Celtic NLP: Training and aligning open models for Welsh language nuances and cultural context.
- Specialised scientific domains: Medicine, healthcare informatics, environmental monitoring, photonics, and engineering simulation.
- Post-training alignment: DPO (Direct Preference Optimization), RLHF, and GRPO reasoning reinforcement.
Autonomous Agents & Multi-Agent Systems
Investigating the next paradigm of artificial intelligence: self-directed agents capable of autonomous planning, multi-step problem solving, code generation, and collaborative deliberation.
- Agent orchestration: Benchmarking multi-agent frameworks (LangGraph, CrewAI, AutoGen) in controlled local environments.
- Tool use & code synthesis: Autonomous pair-programming agents executing in private sandboxes without external API latency.
- Agent alignment & safety: Verifying goal stability, guardrails, and deterministic verification of non-deterministic models.
Edge Intelligence & Embodied AI
Bridging the gap between centralised high-performance clusters and remote, disconnected field environments using our four portable NVIDIA DGX Spark systems (Grace Blackwell, 128 GB unified memory each).
- Offline autonomous agents: Running quantized vision-language models and decision agents directly in the field.
- Real-time sensing & telemetry: Low-latency inference for environmental sensors, marine monitoring, and industrial robotics.
- Quantization benchmarking: Evaluating AWQ, GPTQ, and FP8 precision trade-offs in mission-critical edge applications.
Physics-Informed Neural Networks & Photonics AI
Leveraging Bangor University's world-leading expertise in photonics, optics, and computational engineering to pioneer physical AI and scientific machine learning.
- Optical & photonic simulation: Accelerating electromagnetic simulations, nanophotonics, and optical microsphere modelling.
- PINNs (Physics-Informed Neural Networks): Solving non-linear partial differential equations for thermal and fluid dynamics.
- Inverse design: Machine learning architectures for metamaterial design and optical sensing devices.
Engineered for modern research workflows
Hardware, frameworks, and deployment environments ready for immediate research use.
Central Compute Node (Bangor University)
- System Chassis
- Enterprise 6U Blackwell GPU Server
- GPU Cluster
- 8 × NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB ECC GDDR7 each, 768 GB aggregate)
- Host CPUs
- 2 × Intel Xeon 6530P (64 high-throughput compute cores)
- System Memory
- 1,024 GB (1 TB) DDR5 ECC Server RAM
- Fast Storage
- 60 TB Enterprise PCIe 5.0 NVMe dataset tier
- Interconnect
- High-bandwidth PCIe 5.0 fabric with peer-to-peer memory transfers
Research Software & Framework Stack
- Core Frameworks
- PyTorch 2.x, Hugging Face Transformers, Accelerate, DeepSpeed, FlashAttention-3
- High-Throughput Serving
- vLLM, SGLang, TGI with continuous batching & PagedAttention
- Agent Frameworks
- LangGraph, AutoGen, CrewAI, LlamaIndex, MCP (Model Context Protocol)
- Fine-Tuning Stack
- Unsloth, Axolotl, LLaMA-Factory, PEFT (LoRA/QLoRA), TRL
- Quantization
- AWQ, AutoGPTQ, bitsandbytes, FP8 TensorRT-LLM, GGUF/llama.cpp
- Edge Environments
- 4 × NVIDIA DGX Spark Grace Blackwell units (128 GB unified memory each)
How to collaborate with our research team
Joint Grant Proposals
Partner with Bangor University as Co-Investigators or Project Partners on major research grant bids:
- UKRI & EPSRC funding calls
- Innovate UK Smart & AI programmes
- Horizon Europe & international consortia
- Welsh Government SMART innovation schemes
We can provide dedicated compute resource statements and institutional support letters.
Doctoral & Postdoc Fellowships
Providing compute allocations and isolated project workspaces for qualified PhD researchers and postdocs:
- Thesis experiments requiring multi-GPU fine-tuning
- Independent benchmark validation
- Joint supervision with Bangor faculty
- Summer research internships and visiting scholars
Includes secure scratch directories, isolated containers, and priority batch queue access.
Open Science & Publications
Committed to reproducible research, open-weight artifacts, and high-impact co-authorship:
- Joint submissions to NeurIPS, ICML, ICLR, ACL, and IEEE
- Releasing open LoRA adapters & datasets
- Cross-validation of benchmark claims
- Promoting bilingual and regional AI science
Collaborative publications cite the Welsh Government SMART Capital Equipment Fund.
Interested in collaborating or bidding on a joint grant?
Contact the project lead, Professor Zengbo (James) Wang, to discuss research alignment, compute requirements, or grant partnership.