Blog

Notes on QA that holds up

A man in a call-centre headset with a boom microphone, seen from behind at a desk with one wide monitor. The left part shows the application with one violet bar; the right part shows the AI project assistant in voice mode, with a circle and a waveform, and a dotted line from it to that bar.

Sep 17, 2026 · Engineering

A voice agent for the documentation nobody writes

How we built a voice interviewer to capture project knowledge that never reaches the documentation, and the three constraints it took to make it safe inside a production application: user-scoped authorization, an eighteen-tool read-only surface, and tools that report only outcomes the interface actually produced.

Lukáš KubecDeveloper
Read
A man seen from behind in an office chair at a desk with one monitor showing a checklist of acceptance criteria. Six criteria are pale with empty boxes and one is solid violet with a filled box; a dotted line leads from that one to a small window holding the test built from it.

Sep 1, 2026 · AI Automation · 6 min

It doesn’t work without the tester. That’s the design.

Human oversight of AI mostly fails, and the research says why: accountability alone does not make people check, and showing the reasoning is contested. What is left is the cost of verifying and whether the decision can be skipped. This is the case for treating a tester's approval as the moment a generated test case acquires standing, not as a safety net bolted on at the end.

David KohutProduct Owner
Read
CTA

See it run on your own specification

We’ll run a narrowed part of it live on the call and walk through the output together, and you’ll have the complete generated suite within 24 hours.