Aaron Helwig

Software engineer building AI products
Work

Amira AI voice assistants

Insurance and more, in English and Arabic

Contents
  1. From first conversation to working system
  2. One example: insurance support
  3. Different clients, different integrations

At Amira, I took client projects from the first conversation to a delivered system, most of them voice assistants wired into each team’s own operations: inbound and outbound calls, knowledge drawn from the client’s documents, and actions in their systems during the call. Few of them stopped at a proof of concept.

From first conversation to working system

Each project had its own information, tools and ways of working. I prepared knowledge from documents, organized databases, connected business systems and designed how the assistant should respond, take action or hand a conversation to a person. I carried those decisions through implementation and testing, working through errors with the team.

Across the voice projects, I worked with English and Arabic configurations, context prepared before a call, tools used during it and structured outputs afterwards. The WhatsApp lead work followed the same pattern with property information, CRM records and staff notifications. The channel changed; the need to make the conversation and the underlying operations agree stayed central.

Before
  • Member and policy context, prepared
  • English or Arabic configuration
  • Knowledge cleaned out of documents
During
  • Policy question → member context
  • Nearby provider → facility data
  • Ask where they are, don’t guess
After
  • Structured outcome of the call
  • Handoff a colleague can act on
  • Records that match what was said
The call is the visible part. What surrounds it decides whether it helped.

One example: insurance support

For a large insurance company, I worked on a voice-assistant project involving member records, policy information and provider details. I normalized member data, extracted and cleaned knowledge from documents, and organized facility information so the assistant could retrieve relevant facts during a call.

Two questions sound alike on a call and need different things underneath. “What does my policy cover?” needs the caller’s own member and policy context. “Where is the nearest clinic?” needs the facility data and, crucially, where the caller is right now.

A phone number is not a reliable statement of where someone is, so the assistant asks, then uses the answer to choose a facility or a handoff. Small decisions like that are most of the design: which fact the assistant may assume, which it must ask for, and which it must never guess.

Ask where they are.

Different clients, different integrations

Other engagements called for different combinations of conversation design, knowledge, tools and follow-up. I adapted those pieces to each client rather than treating one successful configuration as a finished solution for everyone. The integrated WhatsApp work for a property company is covered in its own case study.

What I value in this work is seeing a project through as a whole: understanding the need, making the information usable, building the connections and getting the interaction right. The words a customer hears or reads, the records the assistant uses and the information a colleague receives afterwards all have to agree.

Work