Method
The method
Scope, prove, architect, industrialise: the Yelty method for taking an AI project from prototype to production, and the expertise behind it.
AI projects rarely fail because of the model they use. They fail because someone built something that could never hold in production: inaccessible data, impossible integration, unforeseen costs, a security or compliance risk discovered too late.
The Yelty method starts from production and works back to the prototype. Four steps, always in the same order. Each produces a concrete deliverable, and each can be the last if the value is not there.
1. Scope
Goal: pick the right problem.
Understand the business need, the existing technical landscape and the constraints: data, security, regulation, budget. We keep the use case with measurable value and explicitly set the others aside.
Deliverable: a scoping note with the selected use case, its success criteria and the identified risks.
2. Prove
Goal: check that the value exists, on your real data.
A fast demonstrator, built with production constraints in mind. Success criteria are set before starting, so the result is not a matter of impression.
Deliverable: a working demonstrator and an objective measure of its quality.
3. Architect
Goal: design what will hold over time.
Integration with your systems, security, data sovereignty, choice of models and providers, control of inference costs, operability. Every structuring decision is documented and justified.
Deliverable: an architecture document and the related architecture decisions.
4. Industrialise
Goal: go to production and make your teams autonomous.
Deployment pipeline, observability, continuous quality evaluation, operating procedures, then skills transfer.
Deliverable: a system in production, run by your teams.
Principles
- Decide in writing. Structuring choices are recorded as architecture decision records (ADRs): we know why a choice was made, and when to revisit it.
- Measure before industrialising. No go-live without a defined and monitored quality criterion.
- Make costs visible. The running cost of an AI system is designed, not discovered on the invoice.
- Ask the sovereignty question. Hosting, open models or managed services: the choice is made knowingly.
- Pass it on. A mission succeeds when your teams are autonomous at its end.
Areas of expertise
- Enterprise architecture and IT landscape planning — TOGAF, mapping, architecture governance.
- Generative AI in production — chatbots, RAG, agents, evaluation, observability; models from OpenAI, Anthropic and open-weight models.
- Cloud and platforms — Google Cloud, AWS, Azure, Kubernetes, Terraform, CI/CD pipelines.
- APIs and integration — API management, event-driven architectures, microservices.
- Security and compliance — security of AI systems, AI Act readiness.