Agentic pipeline for bordereaux ingestion and checks

An LLM-assisted pipeline that maps messy MGA bordereaux to a standard schema, runs validation checks and drafts the exceptions report for human review.

Design described in general terms. It includes no partner data.

The problem

Every MGA sends bordereaux in its own layout, with renamed columns, merged cells, new fields and changed date formats. Mapping and validating them by hand is slow and error-prone, and the errors flow straight into pricing and monitoring data.

Approach

Schema mapping. An LLM proposes a column mapping from the incoming file to a Pydantic target schema. Mappings are cached per MGA and only re-proposed when the layout changes.
Deterministic checks. All numeric validation is done in plain Python, never by the model: premium reconciliation, date logic, duplicate policies and sum-insured outliers.
Tools, not free text. The agent calls typed tools (exposed over MCP) to load files, run checks and query previous submissions, so every action is logged and reproducible.
Human in the loop. The output is a draft exceptions report and a proposed mapping diff. An analyst approves before anything is loaded.

Design principle

Let the model handle the ambiguous, language-heavy step (what does this column mean?) and leave the arithmetic to code.