Skip to Main Content
Lightning Talk Intermediate MIT License First Talk

Risk-scoring agent-written code before merge: open-source guardrails for the AI pipeline

Proposal status is Review Pending
Session Description

AI agents now write real production code - and data and platform teams own the fallout when generated code is subtly wrong. This talk presents Vouch, an open-source tool that risk-scores AI-written code diffs before merge, and the evaluation discipline behind it: eval-driven flow engineering, house-style and test/lint gates, and in-context domain adaptation of coding models for specific repositories. We will connect this to responsible-AI practice - interpretability of agent decisions, audit trails for compliance regimes like the EU's DORA, and where open-source data infrastructure fits in a governed agentic pipeline. Attendees leave with a concrete checklist for adopting agent-generated code safely in production data systems.

Key Takeaways

How Vouch risk-scores AI-written code diffs before merge; the evaluation discipline behind it - eval-driven flow engineering, house-style and test/lint gates, and in-context domain adaptation for specific repositories; and a concrete checklist for adopting agent-generated code safely in production data systems.

References

Session Categories

Introducing a FOSS project or a new version of a popular project
Engineering practice - productivity, debugging
Talk License: MIT License
Speakers are hidden and anonymized for this event.