Manager control
Feedback submission does not automatically issue an offer. Separate intake, review and the decision to act.
CASE 01 / AI-assisted build
A guest-feedback application, from idea to intake
Defined the workflow, built with AI assistance and opened a deployed guest form.
DELIVERED
Guest intake launched after an AI-assisted build and controlled workflow tests.
EVIDENCE LIMIT
Deployment is verified in the project handoff. Independent manager adoption, live POS redemption and business impact remain unmeasured.
The form was only the beginning.
A feedback form is easy to picture. The harder questions come afterward: who reviews it, what happens if someone submits twice, who can approve an offer and how do we keep track of it? I wanted the whole process to make sense to the manager using it.
What I owned.
I defined the guest journey, the manager’s decisions and the exceptions the application had to handle. I used ChatGPT and Codex to implement and test it, then carried the work through hosted deployment and the guest-intake handoff. The operating requirements and release decisions were mine; the build was AI-assisted.
Where I used judgment.
Exceptions before rollout
Test duplicate submissions, access and state transitions with synthetic records before opening intake.
A bounded launch
Release the guest form while keeping POS redemption, adoption and business benefit as separate checks.
Who had to be in the loop.
I owned the workflow and release decisions, used ChatGPT and Codex for implementation, and coordinated the guest-intake, receipt and protected manager-access handoffs. Staff still needed to verify the real-world transaction and response.
What happens after “submit”?
This small reconstruction makes one operating decision inspectable: intake, review and issuance stay separate. It uses fictional browser-only state, not the deployed application.
SYNTHETIC GUEST / SAMPLE-001
A visit needs a follow-up.
“The order took longer than expected.”
MANAGER REVIEW
No submission yet
Submit the sample to begin. No personal data is collected or sent.
Approval in this example does not issue a real offer. POS verification and reconciliation remain separate.
The decision: a submission creates something to review, not an automatic reward. A repeat submission should not create a second claim.
Inspect the fictional scenarios (JSON) ↓
Actual source check: five isolated tests passed October 25, 2026, covering validation, rating-neutral eligibility, access helpers and safe exports. That verifies those source checks, not live redemption or adoption.
What was actually delivered.
- A deployed guest form, privacy notice and protected manager workflow; intake opened September 28, 2026.
- Synthetic checks covering duplicate handling, rating-neutral review, access, capacity and offer states.
- A receipt handoff and a record of live POS, redemption and adoption checks still to complete.
What remains to demonstrate.
Observe manager use, verify the remaining live redemption path and measure guest participation and operating effort before claiming an improvement.
Building with AI makes the operating questions more obvious. I still have to decide what happens when someone repeats a submission, an offer expires or a manager needs to review something. I find that part of the work especially interesting.