Meet, the AI assistant on WhatsApp

A PropTech business working with Cloud Combinator on AWS. The client is anonymised at their request.

The engagement centred on four focus areas that turn a chat channel into a dependable, production-grade assistant.

Cloud Combinator delivered in five phases over roughly eight weeks, standing up the foundation before layering on the agent, the integrations, and the production hardening.

Success is defined against clear functional and non-functional criteria in the Statement of Work: 99 percent of messages processed successfully, responses maintained across ten or more exchanges, correct host-data retrieval, and a 100 percent escalation success rate, all within a p95 response target under five seconds for text and a 99.5 percent availability floor. The figures below are the engagement's targets and Phase 1 projections.

• Managed agent runtime — Bedrock AgentCore with the Strands framework provides AWS's recommended path for production AI agents, so the client runs a managed runtime rather than bespoke orchestration.
• Ordering guaranteed — A SQS FIFO queue keyed on phone number, with a dead-letter queue, keeps each conversation in sequence and decouples the webhook from AI processing.
• Safe by default — Bedrock Guardrails apply PII redaction, content filtering, and topic blocking, and AgentCore is reached over a PrivateLink VPC endpoint rather than the public internet.
• Personalised over time — Short-term session memory and long-term, user-scoped memory give coherent multi-turn conversations and personalisation across sessions.
• Handover — Terraform modules, CloudWatch dashboards and alarms, structured logs, and knowledge transfer are delivered so the client can run and extend the platform.

Phase 1 is designed with later phases in mind: hybrid search and retrieval, proactive outbound messaging, and multi-channel support are architected for and scoped separately, so is positioned to grow well beyond WhatsApp support as the client takes it on.