LLM REASONING
Case interpretation and advice
Uses the supplied case, retrieved evidence and explicit tool results to reason.
← All systems
AGENTIC UNDERWRITING FLOW
From source document to underwriting context.
A broader research workbench for the machinery around an underwriting decision: domain ontology, company knowledge, document extraction, retrieval, reasoning, rules, pricing and human-visible workflow.
WORKING RESEARCH SYSTEMQUOTE INTAKE
THREE KNOWLEDGE LAYERS
Structure the context before asking for judgement.
The flow separates reusable insurance meaning, organisation-specific evidence and case-specific reasoning rather than compressing them into one prompt.
COMPANY KNOWLEDGE
Documents, policies and precedent
Supports vector retrieval or vector plus knowledge-graph retrieval with precise evidence.
DOMAIN ONTOLOGY
Insurance concepts and relationships
Provides stable vocabulary, extraction targets and connections across source formats.
VISIBLE ORCHESTRATION
Agents and flow are inspectable—not hidden behind chat.
The workbench exposes agent configuration, tool assignment and execution sequence so the workflow can be analysed and changed deliberately.
THE HARD PROBLEM
Many source formats. One target model. No licence to flatten meaning.
AI helps propose structure from documents, tables and narrative. The target schema, ontology, validation rules and review workflow keep that interpretation bounded.
Return precise sentences and paragraphs from company documents with source context.
Use entity and relationship structure where connected evidence improves the answer.
Combine case input, ontology and retrieved company evidence without erasing provenance.
EXPERIMENTAL CONTROL
Compare pipelines without confusing them.
The ingestion system maintains three pipeline families. The optimised and algorithmic variants can each be run under conservative, balanced or high-recall profiles.
PipelineProfilesPurpose
LegacyBaseline
Retained as a reference path for comparison.
OptimisedConservative · Balanced · High recall
Tuned extraction and retrieval behaviour.
Algorithmic labellingConservative · Balanced · High recall
More explicit control over evidence classification.
SYSTEM BOUNDARY
AI proposes. Controlled systems decide what is valid.
01Ontology and schemas define the target meaning.
02Retrieval keeps answers anchored to company evidence.
03Rules and pricing execute outside free-form reasoning.
04People review exceptions and remain accountable.
ARCHITECTURE DISCUSSION
Have a document or knowledge problem worth testing?
Get in touch about source ingestion, ontology design, retrieval quality, workflow orchestration or deterministic rule integration.