David Thomander

Summary

Generative AI Engineer who builds production GenAI applications for insurance, with 10+ years of insurance experience with most workflows in insurance, from sales to underwriting and compliance. 2 years of commercial developer experience. Currently work on advancing generative AI across Nordic initiatives at Tryg.

Skills

GenAI Engineering: Agents & agentic workflows · RAG · structured extraction · tool calling · context engineering · multimodality · model selection & steering

Evaluation & Reliability: Evaluation dataset design · LLM-as-judge · human-vs-AI benchmarking · guardrails & output validation · production monitoring

Technical: Python · OpenAI SDK · LangChain · LiteLLM · Elasticsearch · scikit-learn · Azure AI Services · Streamlit · Git · Docker · Copilot + Copilot Studio ·

Insurance & Business: Private & commercial P&C · claims · underwriting · GenAI strategy & adoption · stakeholder management · technical leadership · teaching & enablement

Experience

• Technical lead for the cross-Nordic GenAI architecture project, setting implementation patterns and delivery standards for 20+ initiatives across three countries.
• Agentic claims and underwriting applications: designed, built and evaluated from discovery to production, 2 now live / 2 in pilot.
• GenAI evaluation, monitoring and governance stack: defined the as-is and target state now used as the reference for new initiatives.
• Advisor to development, data and platform teams on model selection, orchestration, retrieval and agent architecture.

• Production-oriented LLM applications for claims: RAG, structured extraction and agents, with evaluation pipelines built against real cases and human decisions instead of generic benchmarks.
• Technical lead for six student developers over six months, delivering an agentic claims-intake PoC.
• Contributed to enterprise decisions on GenAI architecture, model selection, data access and governance.

• Ten years from outbound sales through customer service, back-office operations and sales training to leading learning & development for 800 employees.
• Deep working knowledge of insurance products and operations, such as claims, sales, backoffice and underwriting.
• Automated my own workflows first, then internal processes and built conversational assistants. This led me into applied GenAI engineering.

Projects

• LLM application that classifies free-text claim descriptions into loss cause and assesses compensability against policy terms; reached a higher accuracy and lower decision time compared to average human claims handler over 11.000 live claims.
• Evaluation harness against ground-truth claims established the baseline for the existing workflow and drove each iteration.
• Migrated from on-prem Llama 3.1 to Azure AI Services via LiteLLM, so models could be swapped without touching application code and to trace usage and spend.

Tech: OpenAI SDK, Elasticsearch, Azure AI Services, scikit-learn, LiteLLM

• Agent that reads an initial claim report, detects missing information and asks the follow-up questions needed to route and process the claim.
• Led six student developers: turned the business problem into architecture, task breakdown and evaluation methodology.
• Raised information completeness by over 100% in early experiments, enabling more accurate routing and fewer follow-up customer contacts. Ran on a locally hosted Llama 2.1 model with 13b parameters.

Tech: Streamlit, context engineering, memory management, evaluation datasets, Llama 2.1

• Agent that reads unstructured broker submissions, extracts underwriting-relevant fields and populates structured pricing inputs in the underwriting workflow, attacking the 30% of time underwriters spend just moving numbers.

Tech: LangChain, Azure AI Services

• Multi-step classifier for claim documents and images combining metadata analysis with visual forensics to flag AI-generated receipts and edited documents. Concept stage.

Tech: LangChain, Azure AI Services

• Multilingual voice assistant answering insurance questions with policy conditions as the source of truth. Built before RAG was a thing, so chunking, indexing and a multi-step grounded prompt chain was custom built. Migrated the bot and our internal knowledgebase to Dynamics365 before i was moved to a new department.

Tech: Llama 2.1, rudimentary RAG, prompt chaining, IBM Watson TTS, Google Translate API

• Project-management assistant that turns meeting transcripts and documentation into plans, briefs, status updates and stakeholder communication from persistent project context.

Tech: Retrieval / embeddings, document ingestion

Earlier personal builds (2021 →): a generative children's-book maker (Eviga Barnböcker), a multi-persona chat app (ChatSquad), a fine-tuned clone of myself trained on ~4,000 of my own messages (MiniMe), an AI-driven Instagram influencer (Peaks'n'Poses) and a Discord bot (Ulf).

Education

• MIT OpenCourseWare, Udemy, a lot of youtube-tutorials and building applications.

• NLP certificate for training rule-based and generative chatbots on the Boost.ai platform.

• Commercial and private P&C insurance.