insurance-llm-guardrails
Guardrail checks and an evaluation harness for insurance LLM assistants.
Python
Guardrail checks and an evaluation harness sitting in front of an insurance assistant, so an unsafe answer fails before a human ever sees it.
Open-source prototypes
Starter and research prototypes – working code exploring agentic patterns for insurance, not packaged products.
mech-gov-claims
Mechanical governance for LLM-assisted insurance claims and underwriting decisions.
Python
Mechanical governance around an LLM-assisted claims decision: the model drafts, and coded rules decide what is allowed to stand.
fair-pricing-audit
Situation testing for discrimination in insurance pricing and underwriting.
Python
Situation testing applied to pricing and underwriting – comparable risks varied one attribute at a time, to expose disparate treatment.
interpretable-tariff
Interpretable one-way relativity tariff model for insurance pricing.
Python
A one-way relativity tariff kept deliberately interpretable, so every pricing factor can be read and checked without running the model.
reserving-ml
Chain-ladder claims reserving (development factors, ultimates, IBNR) with backtesting.
Python
Chain-ladder reserving – development factors, ultimates, IBNR – wired to a backtest, so a change of method is visible before it is trusted.
policy-rag-adapter
Linear query-embedding adapter (triplet loss) to improve policy-document retrieval for RAG.
Python
A linear query-embedding adapter trained with a triplet loss, testing whether policy-document retrieval improves without retraining the model.
stressed-insurance-datasets
Robustness-stress toolkit for insurance ML/LLM systems, with a synthetic motor dataset.
Python
An insurance model stressed against shifted and corrupted inputs on a synthetic motor dataset, rather than judged on a clean public benchmark.
claims-fraud-graph
Synthetic insurance claims-fraud graph generator for graph-based fraud detection.
Python
A synthetic claims-network generator, so graph-based fraud detection can be built and tested without touching real claims data.