Your underwriting knowledge, compiled and compounding
Feed in guidelines, claims data, and underwriter notes. Get a living wiki that writes checklists, calls carrier APIs, and gets smarter with every submission — not a retrieval index that re-discovers the same knowledge on every query.
wiki activity
live
847
wiki pages compiled
73%
faster checklist generation
2.4x
more consistent decisions
Restaurant BOP entity page updated — rooftop signal reinforced
just now
Florida entity page — lint agent flagged +8pt loss ratio
2m ago
Structural load question promoted to standing template
Unlike RAG, the wiki is compiled once and kept current. Cross-references, contradiction flags, and synthesis already exist — no re-discovery on every query.
1
Document ingestion
Guidelines, claims exports, endorsements, and underwriter notes are parsed into clean markdown. Source files are immutable — the LLM reads but never modifies them. A single source may update 10–15 wiki pages on ingest. Azure Document Intelligence
2
Wiki compilation
The LLM ingest agent reads each source and integrates it into entity pages, risk concept pages, and rule pages. Contradictions are flagged automatically. The wiki is a compiled, persistent artifact — not a retrieval index. LLM ingest agent
3
Submission query
When a submission arrives, the query agent retrieves relevant entity, risk, and rule pages. Because synthesis happened at ingest time, the agent reads a structured briefing that already exists. Query agent
4
MCP tool calls
Checklist items requiring live data trigger parallel MCP tool calls: IX carrier API for quote and eligibility, occupancy search for SIC classification, web search for address loss history, and SERFF for state filings. Results pre-populate the checklist. IX MCP · Occupancy · SERFF
5
Write-back
After review, the decision — checklist, MCP outputs, and manual overrides — is filed back as a submission page linked to every entity and risk page it touched. Manual overrides are evaluated for promotion into standing templates. Wiki compounds
Ingest pipelineAzure Document Intelligence
■
PDF / DOCX / email — any format accepted, normalised to markdown
■
Immutable source store — Azure Blob, versioned, never modified
■
Service Bus queue — async ingest job per document
■
One source → 10–15 wiki page updates on average
The wiki
Persistent, interlinked knowledge
Every page is LLM-maintained. Click any page to explore what the wiki knows. Pages are linked — a submission page references every entity and risk page that informed it.
Entity pages
Restaurant BOP
Florida
Liquor liability
Risk concepts
Habitational fire
Cyber — ransomware
Submissions
REST-2026-00441
Live checklist demo
Submission REST-2026-00441
50,000 sq ft restaurant with rooftop bar · Tampa, FL · $4.2M TIV. The checklist was pre-populated from the wiki and parallel MCP tool calls. Check items off as you review, then issue your decision.
Tampa inland — not coastal. Lint agent flagged FL BOP loss ratio +8pts. Refer to manager for TIV >$4M.
Wiki: Florida entity page · lint flag 2026-04-01
SERFF filing — FL BOP rates current?
FL BOP rates filed and approved 2025-11-01. No pending rate filings affecting this risk class.
MCP: SERFF · state filings lookup
Compounding intelligence
The wiki gets smarter with every submission
Drag the slider to see how intelligence compounds as submissions accumulate. Manual overrides become templates. Loss patterns surface in risk concept pages. Appetite signals update in real time.