Research projects exploring knowledge distillation, domain-specific model training, and affordable AI deployment. Led by Dave Gilligan, powered by Blue Note Logic infrastructure, operated through Gilligan Tech ENK in Norway.
A knowledge-distilled 3B parameter model for interpreting Nordic greenhouse gas emissions data. Uses a 32B Qwen teacher model to distill reasoning capabilities into a compact student model, anchored to verified government statistics from SSB and Miljødirektoratet. Designed for CPU-only deployment on AMD EPYC hardware.
A two-stage distillation pipeline producing affordable, Norwegian-specific legal AI. Stage 1: Qwen 27B distils into a 7B "Alpha v0.1" model. Stage 2: Fine-tuning with Norwegian legal intelligence for family law, Barnelova, and Barnevernsloven. Deployed through CorpusAI for Do Better Norge.
Both projects share a core methodology: compress large model intelligence into small, deployable models that organisations can afford to run. This is the innovation thesis — making expert AI accessible without enterprise GPU budgets.
A large model (27B–32B parameters) processes domain-specific data and generates Chain-of-Thought reasoning pairs. The teacher demonstrates how to think about the domain, not just what the answers are.
LoRA fine-tuning transfers the teacher's reasoning into a small student model (3B–7B). The student learns domain-specific reasoning at a fraction of the computational cost, using Unsloth on RTX 5090 hardware.
Organisations upload their own documents (legal briefs, policies, reports) into a private CorpusAI corpus. The model is grounded in verified data — every response is traceable to source documents.
Quantised to GGUF format and deployed via Ollama on commodity CPU hardware. No cloud dependency, no per-query costs, full data sovereignty within EU/EEA borders. Private AI that organisations actually own.
Law firms, NGOs, municipalities — any organisation can fine-tune a model on their own documents. The result is a private AI that knows your domain, not the internet's version of it.
Distilled models run on standard CPU servers. A municipal planning office doesn't need a data centre — they need a model that fits on the hardware they already have.
All training data and inference stays within EU/EEA borders. Hosted on Hetzner infrastructure in Helsinki and Nuremberg. Full GDPR compliance with cryptographic tenant isolation.