Case study · Insurance · Confidential clientProduction AI for the underwriting platform of an international specialty insurer
Backend engineering and AI integrations for a platform processing insurance submissions across multiple products and lines of business. Client kept confidential under NDA.
The project in numbers
Failed extraction records isolated to root cause and fixed in a single incident
People on the distributed team: engineering, QA, design and product
Configuration sources consolidated to enable multi-product onboarding
Of new features behind feature flags, with weekly code freeze and demo
From submission to policy
Broker submissions
heterogeneous documents
AI extraction
AWS Bedrock + validation
Entity resolution
no duplicate clients
Underwriting
quote and bind
The challenge
The platform processes insurance submissions across multiple products. Brokers send documents in heterogeneous formats and every policy requires data extraction and validation — coverage dates, client entities, products like liability or terrorism — before an underwriter can quote and bind.
Manual volume doesn't scale and extraction errors block entire cases. The work was bringing AI to production with the level of control an insurer demands: auditing, permissions, security and predictable releases.
What we built
AI document extraction (AWS Bedrock)
Extraction pipeline with coverage-date validation inside the extraction stage, database status flags and retry logic. Includes resolving a 1,400-record failure incident to root cause.
AI chat over case documents
Natural-language conversation over each case's documentation, demoed to team and stakeholders.
Client entity resolution
Entity matching with rejection statuses and resolution of conflicting signals across data sources.
Broker email automation
Smart triage and automatic follow-up on pending broker responses.
Multi-product onboarding
Architecture and consolidation of 3 configuration sources to bring new lines of business live without rewriting the platform.
Enterprise-grade releases
Mandatory feature flags, weekly code freeze with demo, CI test reporting, secure access via ZPA, private DNS and corporate authentication.
How we worked
A distributed team of 20+ people with dailies, sprints and QA gates. Every merge goes through review, test-environment verification and traceable closure in the ticket system.
The result: a production platform operating multiple products with weekly releases, new lines of business added through configuration, and a passed security assessment with internal access only via ZPA and private DNS zones.
Does your operation need AI in production, not in demos?
Book a call and we'll walk you through how we take AI systems to production with the control a regulated industry demands.
Book a call