Aaron Helwig

Software engineer building AI products
Work

Amira AI lead assistants

WhatsApp lead handling for real estate

Contents
  1. Two conversations, one channel
  2. Keeping replies grounded
  3. Making changes without losing the thread

At Amira, I worked on a WhatsApp lead-handling system for a real-estate team. It joined customer conversations, property context, staff notifications and follow-up records. The central problem was ownership: automation should help a lead move forward without competing with the person already handling it.

Two conversations, one channel

Customer messages and staff replies needed different interpretations. A customer asking about a viewing should enter the conversation flow; a staff member responding to an assignment should update the handoff. I designed the routing and logging around that distinction, including the interval in which an assigned person could respond before assistant fallback.

The words used for staff actions mattered. A brief acknowledgement such as ‘OK’ does not establish that someone contacted a customer. I worked on explicit claim, defer and contact signals so the recorded state reflected an actual action rather than an optimistic interpretation.

Staff replyWhat gets recorded
Claiming this oneAgent owns the lead. The assistant steps back.
Later todayDeferred. The assistant keeps the customer company.
Called them ✔Recorded as contacted, with who and when.
OKNothing. An acknowledgement is not an action.
(no reply)Window ends: the assistant follows up, staff are told.
Each staff reply either changes who owns the lead, or deliberately changes nothing.

Keeping replies grounded

I connected listing and CRM context to the assistant’s responses and worked through failures involving mismatched listings, incomplete property information and requests for viewings. The assistant needed to distinguish an enquiry from a confirmed appointment, and a requested handoff from one that had actually happened.

A request is not a confirmation. That distinction belongs in both the conversation and the records used by the team.

Making changes without losing the thread

The work grew to include workflow audits, database and event-log structure, provider behavior, and controlled release packages. Repeated messages, interrupted actions and uncertain delivery results needed deliberate treatment: retrying blindly could duplicate the very notification the team was waiting for.

I kept the later redesign separate from the existing operational workflows while testing its import and runtime behavior. It did not reach a production cutover, but testing showed exactly where the platform departed from the workflow specification, which gave the next iteration a concrete list of problems to resolve.

Work