An agentic text-to-order system built for one of the largest QSR chains in the United States. Customers could place orders, customize items, and complete payment entirely through SMS — no app download, no account creation, no friction. The payment flow ran through an iOS app clip triggered from the conversation, creating a seamless end-to-end experience that never left the messaging thread.
The System
Two builds were developed across the engagement. The first was NLP-heavy and built on Amazon Bedrock. The second — and more advanced — was purely agentic, using specialized agents for distinct responsibilities: intent organization, cart assembly, and rapid response while longer workflows completed in the background. This architecture allowed the system to handle conversational complexity without sacrificing response time.
The iOS payment flow was a particular technical achievement. Users could initiate an order over SMS, receive a payment link that launched an app clip inline, complete checkout, and return to the conversation — all without leaving Messages or downloading anything.
My Role
I led QA across both builds, with primary responsibility for the second agentic system. This included:
- Managing over 1,000 test cases covering order accuracy, conversational accuracy, edge cases, and failure modes.
- Test case design that included intentional typos, highly complex multi-item orders with extensive customizations, unavailable menu items, and ambiguous user intents.
- Engineering automated testing workflows in Power Automate during earlier builds before a purpose-built testing tool was developed by the engineering team.
- Maintaining the test suite and producing regular accuracy reports once automated tooling was in place.
- Agile project management responsibilities: sprint planning, scrum facilitation, and retrospectives.
- Ongoing product research to inform system improvements and roadmap decisions.
Outcome
The second build achieved over 85% order accuracy across a test set of 500+ cases on multiple occasions — a meaningful benchmark for a fully conversational, agentic ordering system operating at the complexity level a major QSR menu demands. The project was ultimately sunset due to business factors outside the technology's performance.