The insurer ran a mature but end-of-life virtual agent handling policy servicing, premium queries and claims intake. The mandate: migrate to an AWS-native stack without disrupting a live book of business.

We inventoried the legacy estate, designed a shared-intent Lex V2 architecture with a Lambda orchestration layer into policy administration APIs and a Drools rule engine, and rebuilt flows in modular packages sized to fit Lex quotas and release cadence.

Delivery ran as phased releases with parallel-run verification: each wave shipped through dev, SIT and UAT gates in a Terraform-managed pipeline before production traffic shifted. Conversational regression suites caught intent drift before customers did.

The result is an estate the client's own engineers now extend — with a release cycle measured in weeks, and a platform ready for Bedrock-powered retrieval on top.

Customer service in the Gulf is bilingual by necessity — and Arabic support is where most bot programs quietly fail. The client needed both languages at equal quality across chat and voice.

We built a single conversation architecture with per-locale Lex bots sharing an orchestration layer, dialectal utterance tuning for Gulf Arabic, and Bedrock retrieval grounded on the insurer's own bilingual policy documents with citation-backed answers.

Guardrails enforce PII masking and keep commercial questions with humans. Every generated answer carries its source, which is what made the compliance sign-off straightforward.

Self-service containment cleared sixty percent within the first quarter, with Arabic sessions performing on par with English — the design goal that mattered most.

The group's contact center changes moved through email approvals and console clicks — unrepeatable, unauditable, and slow. The engagement goal was engineering discipline, not new features.

We codified the entire Connect and Lex estate in Terraform modules, established trunk-based branching with plan-review gates, and wired SonarQube and security scanning into a Jenkins-and-GitHub pipeline serving distributed teams.

Environment provisioning that took weeks of tickets became a pipeline run. Audit questions became git history.

The delivery model now serves multiple business units from one module library — the reuse economics that make a platform team pay for itself.