The Questions Underneath AI.
Advice, evaluation and models from a working AI researcher. Sebastian Vaderaa leads the AI practice and publishes through Soma Labs.
A model of what happens next.
A small model watches a moving point and predicts where it goes next. Blue is the prediction, white is where it went.
How I can help
- Advisory
What AI can and cannot do for your organisation, judged from how the systems actually work rather than from vendor claims. Build, buy or wait.
- Evaluation and testing
Blind, pre-registered tests of whether a model or AI system does what it claims, against a proper control.
- Systems and agents
Architecture for retrieval, agents and tool-using systems, designed so that their behaviour can be tested.
- Research and models
Research partnerships and custom models where the problem genuinely needs one, including world models.
Where Vaderaa is going
- TodayArchitecture and governance
Twenty-five years of cloud and enterprise architecture on national programmes, and the review boards that keep them governed.
- Now addingAI governance
The same discipline applied to AI in regulated industries, led by Kabir Vaderaa.
- NextAI research and evaluation
Research-grade advice, testing and models, led by Sebastian Vaderaa through Soma Labs.
Sebastian VaderaaAI lead, Soma Labs
Sebastian Vaderaa leads the AI research and engineering. He builds frameworks for world models and deterministic marking, published through Soma Labs.
Soma Labs works on the questions underneath AI rather than the products on top of it, and every claim is pre-registered and tested blind.
Talk to SebastianThe work
Can a system build its own world?
The programme →Can an essay be marked without being understood?
The study →AI research and guides
- AI FundamentalsRead
Core concepts: tokenisation, weights, context windows, and transformer architectures explained clearly for technology leaders.
- Open-Weight AIRead
Evaluating Llama, Mistral, and Qwen open-weight models for self-hosted data privacy and cost optimization vs API providers.
- Agentic AIRead
How autonomous, tool-using AI systems actually work, what they can and cannot do, and what governing them requires in regulated environments.
- Model MatrixRead
Independent comparison matrix across flagship open-weight and commercial LLMs covering context, cost, and specialties.
Publications from Soma Labs
AI governance for regulated industries is led by Kabir Vaderaa, alongside the architecture practice.
AI governance →Bring Sebastian a hard AI question.
Advice, evaluation or research. Start with a short note on what you are trying to do.