AI 401 · Year 4 · Semester 1 · 3 credits · AI & Machine Learning

NLP & Large Language Models for Insurance

Transforming unstructured insurance text into auditable actuarial intelligence—from policy wordings to claims triage.

Season 1 · 8 episodes

Latent Coverage: The NLP Reserving Engine

When catastrophic storm backlogs and unstructured claims notes threaten a commercial insurer with insolvency and regulatory sanctions, lead actuarial data scientist Maya Lin must engineer an end-to-end, mathematically governed LLM pipeline to triage fifty thousand files before the quarterly board audit.

Protagonist · Maya Lin, Lead Actuarial Data Scientist & NLP Specialist at Apex Mutual

Setting · Apex Mutual's actuarial headquarters in Chicago, Illinois

Stakes · A 62-million-dollar loss adjustment expense surge, an imminent AM Best solvency downgrade, and multi-state bad-faith regulatory audits under ASOP 23 and ASOP 56.

Pick an episode

Episodes

Season finale: Maya deploys a fully governed, privacy-preserving hybrid RAG and constrained extraction claims triage engine that processes 12,000 live catastrophe claims in 45 minutes, slashing processing cycle time from 18 days to four minutes with zero schema failures.

From the lecturer: Welcome to AI 401. Unstructured text represents over eighty percent of all insurance operational data—from hand-written adjusters' logs to five-hundred-page commercial policy wordings. As actuaries and AI engineers, treating text as a black box is unacceptable. You need a mechanistic understanding of how tokens flow through transformer attention heads, how vector spaces represent coverage semantics, and how retrieval architectures ensure absolute factual fidelity. To master this course, keep two notebooks open. In your mathematical notebook, trace matrix dimensions, compute scaled dot-product attention by hand, and sketch information retrieval pipelines. In your coding notebook, implement every architecture using PyTorch and Hugging Face, from subword tokenisers to quantized LoRA fine-tuning scripts. While AI 401 does not map to a traditional preliminary exam, it directly operationalises the modern data science directives of the Casualty Actuarial Society, the Society of Actuaries predictive modeling guidance, and Actuarial Standard of Practice Number 56 on Modeling. Regulators demand auditable, privacy-compliant AI systems. Approach every architecture not just with an eye for loss minimisation, but with the professional discipline required to deploy mission-critical financial workflows.

Cold open

Maya arrives at her desk at six in the morning to find the claims floor in crisis. Forty-two thousand first notice of loss notes from a catastrophic midwest hail and wind storm are piled up as free text, and the legacy bag-of-words model just classified three thousand severe commercial roof collapses as routine glass claims because of misspellings and out-of-vocabulary jargon.

preparing lecture…

Transcript

Contrast rule-based, word-level, and subword (Byte Pair Encoding) tokenisation schemes on unstructured insurance text.
Derive the embedding lookup mechanism and formulate the Skip-Gram with Negative Sampling objective.
Compute cosine similarity and document-level pooled embeddings to quantify semantic distance between adjuster notes.
Evaluate embedding representations under Actuarial Standards of Practice regarding data quality and model governance.