Four ways in. All of them start with a conversation.

AI strategy & advisory

The most expensive AI mistakes are made before any code is written. I work with leadership and engineering together to separate the use cases that will pay for themselves from the ones that will consume a quarter and produce a slide.

I've done this as an internal technical advisor for AI product procurement, POC initiatives and cross-department integrations — so the recommendation accounts for security, compliance and the team you actually have.

• Prioritised use-case map with effort and risk
• Build vs. buy recommendation per capability
• Vendor and model shortlist with rationale
• 90-day sequenced delivery plan

Agent & LLM product build

Hands-on delivery: assistants, retrieval systems, agents with real tool access, and the automations around them. Python and FastAPI on the backend, Next.js or React where there's a UI, LangChain or Haystack where it earns its place.

Previously: a generative AI assistant platform and LLM proxy for a Nordic bank's developers, an internal enterprise ChatGPT with proper authentication, and agents plus MCP servers at a consumer health product.

• Working system in your cloud, not a notebook
• Model gateway / proxy with access control
• Retrieval and vector search layer
• Handover to your engineers, documented

Evaluation & LLM ops

If you can't measure your AI feature, you can't improve it and you can't defend it. I build the evaluation layer: LLM-judge frameworks, curated test sets, regression runs wired into your pipelines and dashboards.

Built on Databricks with Airflow orchestration in a previous role, and on plain CI where that's the right size for the team.

• Eval suite with agreed quality metrics
• Automated regression on every prompt change
• Observability, tracing and cost visibility
• Release criteria your team can hold