Development PracticesLightning talk20min
How AI Helped Us Catch Regressions in Code With Zero Coverage
The talk presents a lightweight testing approach for fast-growing codebases: reusable flyweight test scaffolds combined with AI-generated test cases from code changes. It includes a live demo catching a regression, plus lessons learned about AI failures such as hallucinated assertions and false confidence.
talk.summaryAiDisclaimer
Youssef AmeachaqIdemia Public Security
A few months ago, I joined a project that had grown fast. New features every sprint, but almost no tests. Every change felt risky. We were scared to touch old code because we never knew what would break.
One day, a small fix broke a feature in production. Nobody noticed until a customer complained. That was the moment we said: we need a safety net, but we don't have time to write hundreds of tests from scratch.
So we tried something different. Instead of writing full, heavy tests for everything, we used a flyweight approach: small, shared test scaffolds that we reuse across many cases. And we used AI to help generate the specific test cases on top of that scaffold, directly from our code changes.
In this talk, I will show you exactly what we did. I will do a live demo: take a real code change, generate a flyweight test with AI in front of you, and show how it catches a real regression. I will also show you where AI failed us. It hallucinated assertions. It gave false confidence sometimes. I will explain how we caught that and what we changed in our process.
One day, a small fix broke a feature in production. Nobody noticed until a customer complained. That was the moment we said: we need a safety net, but we don't have time to write hundreds of tests from scratch.
So we tried something different. Instead of writing full, heavy tests for everything, we used a flyweight approach: small, shared test scaffolds that we reuse across many cases. And we used AI to help generate the specific test cases on top of that scaffold, directly from our code changes.
In this talk, I will show you exactly what we did. I will do a live demo: take a real code change, generate a flyweight test with AI in front of you, and show how it catches a real regression. I will also show you where AI failed us. It hallucinated assertions. It gave false confidence sometimes. I will explain how we caught that and what we changed in our process.
Youssef Ameachaq
Senior Technical Lead with over 10 years of experience in software engineering, specializing in biometrics and digital identity solutions. Throughout my career, I have worked with global organizations such as Oracle and Orange Business, as well as innovative startups and growing companies.
I have extensive experience leading cross-functional teams, driving complex projects from conception to delivery, and providing technical direction to ensure scalable, secure, and high-quality solutions. My expertise spans software architecture, engineering leadership, project execution, and aligning technology initiatives with business objectives.
I have extensive experience leading cross-functional teams, driving complex projects from conception to delivery, and providing technical direction to ensure scalable, secure, and high-quality solutions. My expertise spans software architecture, engineering leadership, project execution, and aligning technology initiatives with business objectives.