Agentic systems for insurance risk

Multi-step agents on Bedrock, retrieval over insurance data, and decision APIs that actuarial teams can actually call.

2025 — now · current

99.9%

Target SLO · risk-scoring workflows

Decision APIs

Surface

The situation

Enterprise AI at an insurer only counts if it survives an eval harness and lands as a production API. The work is not a chatbot. It is agents that reason in steps, retrieve from messy insurance records, and hand a structured answer to a risk workflow that already has an SLO.

The work

I architected agentic systems on AWS Bedrock: multi-step reasoning, tool calling, structured output schemas, and offline eval harnesses.
I built RAG pipelines over structured and unstructured insurance data — embedding selection, retrieval tuning, and hallucination guardrails.
I led cloud-native platform work on Lambda, Step Functions, DynamoDB, and SQS, aimed at a 99.9% SLO across event-driven risk-scoring workflows.
I worked with actuarial and data science so model output becomes a decision API, not a notebook.