An advisory AI code-review agent for GitLab CI, built to find high-signal issues before merge while keeping humans in control.
Role
Independent Designer & Engineer
Year
2026
Stack
Go · PostgreSQL · GitLab CI · Goose CLI · Prometheus · Grafana · Kubernetes
Team
Tapsi Developer Platform
MergeScout is an advisory AI code-review agent for Tapsi’s GitLab CI workflow. I identified the review-quality problem, framed the product, designed and implemented the service, ran the pilot, and completed the rollout independently.
Problem
Code review needs coverage for security issues, unsafe deployments, convention violations, and edge cases, without turning every merge request into a stream of low-signal bot comments or unbounded model cost.
Approach
Repository-aware agent: The agent explores repository files and Git history with Goose CLI, follows project-specific instructions and skills, and returns structured line-level findings.
Advisory by design: MergeScout never blocks a merge. Engineers keep the decision, while low-confidence findings are filtered before publication.
Evaluation and feedback: A weekly feedback loop compares findings with human review discussions using an LLM judge and records reviewer reactions as calibration signals.
Cost controls and operations: Per-project budgets, token monitoring, Prometheus metrics, Grafana dashboards, and Kubernetes deployment keep the service observable and bounded.
Adoption
MergeScout is adopted across 50+ projects, each averaging 15 reviewed merge requests per week. It is designed to surface useful issues without replacing human review.
