Debugging Methodology and Guides
AI Debugging Methodology Inspired by Medical Diagnostics
-
Introduction
Purpose of this repo: share a structured debugging framework drawn from emergency medicine practices Target audience: AI/ML practitioners of all levels, especially those without clinical background Why medical diagnostics? High-stakes, time-sensitive, iterative decision-making under uncertainty
Core Principles from Emergency Medicine
Rapid Assessment & Triage
Briefly survey symptoms (error messages, logs)
Prioritize critical failures (data pipeline breaks vs. minor warnings)
Hypothesis Generation
Formulate multiple potential causes quickly
Use differential diagnoses (e.g., "Did the tokenizer crash or the attention mask fill?")
Focused Testing
Run focused checks (unit tests, small data subsets) to confirm or rule out causes
Iterative Plan A → B → C…
Develop successive remediation steps when initial fixes fail
Continuous Monitoring
Track metrics (loss curves, latency) to detect emerging issues early
Team Communication & Handoff
Clear notes/logs for collaborators or future you