Code review API your agent pipeline calls on every pull request

Coding agents now open pull requests under an engineer's own account, a dozen inside an hour when one engineer queues a dozen tasks before standup. Seat-priced reviewers meter that engineer by the hour: one entry plan reviews five, and the rest wait for the next hour or bill $0.25 per file. The human reviewer, already the team's slowest step, opens the other seven diffs cold.

Checkrail Tecnologia Ltda. was registered in Curitiba, Brazil, in January 2022 to rank static-analysis warnings by whether reviewers had acted on similar ones before. When agents began opening pull requests in 2025, that ranker became the part of Checkrail deciding which findings get posted. The pipeline calls one endpoint when a draft opens, passing the agent's task, and gets back pull request comments and a verdict: back to the agent, ready for a person, or a person first.

Judging whether a diff does what its ticket asked is work no rule can do, and Checkrail spends its GPU time there. Four fixed checks take the mechanical part: secrets, removed tests, CI and dependency changes, paths outside the task's code owners. Intent and breakage go to an open-weight code model. It reads each hunk with its whole file, the callers an embedding index finds and the ticket, and proposes findings with line ranges and suggested patches.

A finding reaches the pull request only at 0.80 confidence or above, and Checkrail never approves, blocks or merges. Findings between 0.50 and 0.80 return in the JSON for the pipeline to show or drop; lower ones are discarded. Approval stays with the reviewer the code-owners file names. Each comment that reviewer resolves or dismisses becomes a labeled pair, finding, hunk and verdict, retraining the account's reranker overnight. No customer code trained today's code model.

Review load arrives in waves set by when engineers dispatch their agents. One person's dozen pull requests land before standup, another wave follows lunch, and nights are nearly empty. A new repository adds a one-off backfill that embeds every file at its default branch. Many teams' security policies forbid sending source code to an outside model provider, so every model runs on cloud GPU capacity we control in Frankfurt, beside diffs held on cloud infrastructure in Frankfurt.

The prompted code model behind every review runs today on hourly-billed GPU instances in Frankfurt, each started from zero when a wave begins. By March 2027 a smaller model replaces it, fine-tuned only on pairs from accounts whose owners opt in, and small enough to serve a review from one GPU instead of two. By September 2027 a baseline of reserved GPU capacity in Frankfurt keeps every repository's index warm, so a morning wave no longer starts cold.