LLM Safety as a Quality Gate
A GitHub App that enforces LLM safety standards at the PR level
the same way linters enforce code quality.
While LLM evaluation frameworks (EleutherAI Harness, DeepEval, HELM) and ML fairness libraries (AIF360, RAI Toolbox) exist independently, no prior work integrates automated bias/robustness evaluation as a native GitHub App with PR-level merge enforcement, confidence scoring, and longitudinal audit logging.
Kavita Jadhav
K11 Software Solutions LLC · Texas, United States
LLM failures go undetected until production
Demographic bias in training data — invisible to functional tests
Imbalanced accuracy across groups — missed at code review
Brittleness under adversarial inputs — surfaces only at scale
⚠ No existing tool integrates LLM safety into PR review as a blocking merge gate
Shift-left LLM safety into every pull request
Like linters enforce code quality — LLM Eval Agent enforces safety standards at merge time
6 engineering innovations
3 architectures · 5 test categories
Gate thresholds: Bias ≥ 80% · Fairness ≥ 80% · Robustness ≥ 75%
Fairness: 0% across ALL architectures
This is not an architecture defect. All three models were fine-tuned on SST-2, which contains gender-correlated sentiment patterns. Remediation requires dataset-level intervention — not model swapping.
✓ Surfaced automatically on the first PR — exactly the shift-left outcome intended
End-to-end in 21 seconds
Available now as an installable GitHub App
Installs on any repository. No infrastructure to manage. Safety evaluation automatically on every PR.
Also available on Zenodo · ResearchGate