Software teams have never moved faster. With AI-powered tools, developers can now generate code, tests, and solutions in seconds — dramatically increasing productivity and accelerating delivery.

But while speed has evolved, quality hasn’t kept up.

This creates a dangerous illusion: the belief that more output means better outcomes.

In many organizations, what is perceived as a strong quality culture is often just a collection of tools, processes, and checklists — a “quality façade.” And with the rise of AI, this façade becomes even more convincing, as automatically generated code and tests give a false sense of confidence.

In this talk, Joana Silva explores what true quality culture really means and why it’s more critical than ever in an AI-driven world. She will challenge common assumptions about quality, highlight the risks of over-reliance on AI, and explain how speed can unintentionally increase technical debt, defects, and loss of trust.

Through practical insights and real-world experience, this session will cover how organizations can move beyond the illusion of quality and build a culture where quality is truly owned by everyone.

Because in the age of AI, one thing becomes clear:
AI doesn’t replace quality culture — it amplifies it.The software engineering world is currently trapped in the “Happy Path Fallacy” when it comes to AI. We build an agent, test a few isolated prompts in a playground, verify the “vibes,” and push to production. But traditional testing methodologies break down with AI. Agents are non-deterministic, they take multi-step cognitive trajectories, and they suffer from silent drift when the underlying models update. For engineering and QA teams, “vibe checks” do not scale.

To build reliable AI systems, we need to transition from the dark art of ad-hoc prompt crafting into a rigorous engineering discipline. This session is designed for technical leaders and engineers who need to bring predictability to unpredictable models. We will explore how to build automated evaluation pipelines that act as the ultimate CI/CD gatekeeper for GenAI applications, treating evaluations as the immutable product specification.

We will dive deep into a practical Eval Toolkit comprising four scalable patterns: Deterministic Assertions for strict formatting, LLM-as-a-Judge for open-ended quality (and how to overcome its inherent biases), Trajectory Evaluations for workflow efficiency, and Adversarial Evals for stress-testing edge cases.

By anchoring these metrics in comprehensive Tracing (using tools like MLflow), teams can finally embrace “Eval-Driven Development” (EDD). This talk provides a concrete, code-backed approach to stop guessing, catch regressions automatically, and turn real-world production failures into permanent test cases.