p055ible
Get in touch

Case study

LLM-based config verification at scale

Built an automated verifier that checks interdependent JSON configs across GitLab repos, reducing manual validation time and release risk.

2024-10

Solutions Architect

ML Agents Automation DevOps
Manual verification reduced from hours to minutes
Cross-file inconsistencies detected before release
Repeatable verification for non-engineering teams
GitLab Python Vector Search LLM
LLM config verification flow diagram

Challenge

Large, deeply nested JSON configs lived across multiple GitLab repos and branches. Changes in one file often required updates elsewhere, but manual verification was slow, brittle, and missed cross-file dependencies.

Approach

  • Built a web tool that indexes JSON key paths and maps them to reference docs.
  • Added LLM-guided verification with structured outputs (PASS/FAIL/UNKNOWN).
  • Implemented cross-file consistency checks and optional diff coverage.
  • Enabled MR-only patch proposals with human review.

Results

  • Verification became repeatable and accessible for non-engineering teams.
  • Cross-file inconsistencies surfaced before release checks.
  • Manual review time dropped from hours to minutes.

Key metrics

Measured impact
MetricBeforeAfter
Verification timeHoursMinutes
Cross-file checksManual spot checksAutomated consistency rules
AuditabilityAd hoc notesStructured PASS/FAIL reports
Implementation notes
  1. Mappings provided the best balance between automation and precision.
  2. Deterministic outputs improved trust and adoption.
  3. Auto-proposals helped, but merge decisions stayed with humans.