High Signal: Data Science | Career | AI Episode 36: AI and the Judgment Problem in Data Science
22 snips
Mar 19, 2026 Andrés Bucchi, a senior data leader at LATAM Airlines focused on scaling data teams and secure AI deployment, and Dawn Woodard, a distinguished engineer with experience building experimentation and analytics platforms, discuss AI’s impact on analytics. They cover the source-of-truth challenge, verifiable outputs and catalogs, agent-first tooling, experimentation bottlenecks, and how roles shift toward validation and trustworthy AI workflows.
AI Snips
Chapters
Transcript
Episode notes
Conversational Queries Require Canonical Data Definitions
- Conversational querying democratizes analysis but amplifies the need for clear source-of-truth definitions for ambiguous concepts like user sessions.
- Dawn built a quick vibe-coded app, yet at Uber four different "user session" definitions showed why annotation and canonical sources matter.
Build A Strict Catalog Before Broad AI Use
- Prioritize high-quality data cataloging and strict documentation before broad AI rollout to make AI outputs verifiable.
- Andrés focuses AI on easily verifiable tasks first (legal processes, code, hypothesis analyses) to funnel testable ideas downstream.
Use Trusted Evals To Keep Agents On Track
- Implement evals and trusted reference points to give LLM agents offline and live checks against known numbers.
- Jeremy recommends using both offline metrics and live gut-checks so agents can double-check answers.