Why it exists
A green test run can hide a broken checkout or an assistant that only answers correctly some of the time. A release decision needs evidence that covers both.
Portfolio / CASE STUDY
Public engineering lab
Release Evidence Lab.
A synthetic storefront and release gate that use fault injection, browser tests, and repeated AI evaluations to explain what can ship.
MY ROLE
Independent test and release engineering
What I built
I built a small storefront, a risk-based test runner, and a policy that emits SHIP, REVIEW, or BLOCK. A browser replay lets visitors remove evidence or introduce faults and see how the verdict changes.
Project in view
Source for this capture
Engineering decisions
Make the gate deterministic
Required checks and evidence drive the release verdict. Model suggestions run in shadow mode and can propose additional checks for review.
Repeat and score assistant cases
Golden prompts exercise policy boundaries across repeated runs. A separate benchmark compares the model judge with authored reference labels.
Inject failures on purpose
Recorded scenarios distinguish selector drift from checkout regressions, duplicate charges, privacy leaks, and assistant regressions. The reports show which check caught each fault.
Capabilities in practice
- Release policy design
- Browser and API automation
- Model evaluation
- Reproducible fault injection
Status & scope
Independent lab with a fictional storefront and recorded evidence. The hosted demo replays the release scenarios.