QA Copilot
AI-powered QA system.
The problem
Manual QA is slow and inconsistent, and automated end-to-end tests are famously flaky — eroding trust until teams ignore failures and ship bugs anyway. QA becomes a bottleneck instead of a safety net.
Context
Having built browser automation for Snappit, the leap to QA was natural: the same resilience techniques that make automation reliable can make tests reliable. QA Copilot applies AI agents to the testing lifecycle itself.
Architecture
QA Copilot observes application behaviour to generate meaningful tests, uses AI to triage failures (real bug vs. flake), and self-heals brittle selectors. It plugs into CI so the safety net runs continuously.
- Behaviour-driven test generation.
- AI failure triage that separates signal from noise.
- Self-healing selectors built on the Snappit recovery layer.
- CI-native so confidence is continuous, not a gate.
Technical challenges
Trust
Teams have to trust the copilot’s judgement on what failed and why. Calibrating confidence and being transparent about reasoning is the whole game.
Flake vs. bug
Distinguishing a genuine regression from environmental flakiness is hard. AI triage plus historical signal makes the call defensible.
Engineering decisions
Reuse the Snappit recovery layer
Self-healing selectors were a solved problem from Snappit — composing my own platforms compounded the value.
AI assists, humans decide
The copilot proposes; engineers approve. That keeps it useful without overstepping trust.
Technologies
Results
Faster releases with higher confidence, and far less time lost to brittle test maintenance — QA that scales with the product instead of fighting it.
Lessons learned
- Reusing your own platforms compounds leverage.
- AI tools win on trust, earned through transparency.
What I’d improve today
- Risk-based test prioritisation from production telemetry.
- Auto-generated reproduction steps attached to every flagged failure.