BHSI Risk Classification System

Directors and officers underwriting required analysts to gather financial, legal, and market context from multiple sources before making a risk recommendation.

The workflow was slow, inconsistent across reviewers, and difficult to scale when submission volume increased.

The system combines agent-based research, retrieval over structured and unstructured data, and hybrid rules for risk scoring so model output remains grounded in underwriting logic.

Instead of generating a single free-form summary, the pipeline produces explicit risk factors, supporting evidence, and a normalized recommendation payload.

FastAPI coordinates the underwriting workflow while background tasks fan out research requests across financial, news, and document sources.

Gemini-powered reasoning sits behind guardrailed prompts and structured schemas so downstream scoring and UI layers can consume predictable fields.

The project turned a multi-hour analyst workflow into a minutes-long assisted review, making the platform useful as a first-pass risk triage system.

The biggest product win was not just speed, but producing an output shape that underwriting teams could review and challenge instead of treating the model as a black box.

Insurance workflows need explainability as much as model quality. Structured evidence and rule visibility matter more than polished prose.

For production-readiness, the next step would be stronger observability around agent failures, retrieval quality, and override behavior from human reviewers.