For teams shipping RAG & agent features on GitHub
Know whether an AI change is safe to merge.
Vouqis Verify turns every AI pull request into a clear release decision — what changed, which tests and evaluations passed, what evidence is missing, and what requires human approval.
GitHub-first · Advisory mode first · No production access required
The problem
AI diffs are not just code diffs.
“Passing CI is not proof that your AI feature is safe.”
Behavior changes without a visible diff
A prompt, model setting, or retrieval config changes by a few lines. The diff looks small. The model's behavior in production can shift far more than the diff suggests.
Evidence is fragmented
What changed lives in code. Whether it's safe lives across tests, evaluations, CI runs, and traces — in different tools, updated at different times.
Reviewers approve on partial information
Without a consolidated view, a reviewer either reconstructs the evidence by hand or approves on trust. Neither scales as AI surface area grows.
One PR. One evidence pack. A clearer decision.
- Open the diff and guess which files are AI-related
- Search Slack or CI history for the last eval run
- Open the eval dashboard in a separate tab
- Cross-reference commits, prompt versions, and model IDs by hand
- Approve on a hunch, or let the PR sit
- Changed AI surface listed automatically, by category
- Behavioral and evaluation differences summarized
- CI, test, and evaluation results in one place
- Missing checks and policy failures flagged
- A clear merge-risk recommendation, posted on the PR
How it works
From diff to decision, in three steps.
- 01
Change
A pull request changes AI-related code — prompts, model settings, retrieval, tool permissions, or agent logic.
- 02
Evaluate
Vouqis Verify detects the change and runs your existing evaluation command against base and head.
- 03
Review
A structured evidence pack is posted to the PR, with a merge-risk recommendation.
Policy
Make “safe to merge” a policy — not a feeling.
- Prompt changedRequire prompt-regression eval + reviewer approval.
- Retrieval access changedRequire authorization test + retrieval-quality eval.
- Tool permission changedRequire schema validation + security approval.
- Model version changedRequire quality, latency, cost, and structured-output checks.
What you get
Everything a reviewer needs before approving.
- 01
AI surface changed
Prompts, model settings, retrieval config, tool permissions, or agent workflow — identified automatically.
- 02
Behavioral and eval differences
How the model's outputs shifted between base and head, on the scenarios that matter.
- 03
CI, test, and evaluation results
Pulled from the checks your pipeline already runs — no separate dashboard to open.
- 04
Missing checks and policy failures
Flags when a required evaluation, like a safety check, didn't run at all.
- 05
Representative traces
Concrete before/after examples of model behavior, not just aggregate scores.
- 06
A clear merge-risk recommendation
A written rationale and a recommendation — the reviewer still decides.
Security & integrations
Runs the evaluation tooling you already have.
Vouqis Verify does not ask your team to adopt a new evaluation framework. If your checks already run in one of these, Verify calls the same command against base and head.
- GitHub App scoped to the repositories you choose, with read-only access where supported by the current integration.
- Your code and evaluation output are used only to build the review package — not for training.
- Verify never merges code. It posts a review package; a person decides.
FAQ
Before you write in, the short answers.
- Does Verify replace reviewers?
- No. Verify assembles the evidence: changed AI surface, eval results, and rationale. A person still decides whether to merge.
- Does Verify replace our evaluation framework?
- No. Verify runs the evaluation command your team already uses, whether that is pytest, Promptfoo, Braintrust, or a custom script. It does not add a new one.
- Will it block merges?
- No. Verify posts a review package on the PR. It does not gate the merge button, and it never merges code itself.
- What does installing it touch?
- A GitHub App or Action scoped to the repositories you choose. It reads code and evaluation output to build the review package.
Founding Design Partner Program
For GitHub-based AI teams shipping RAG or agent features.
- A 30-minute AI PR verification workflow audit
- A tailored verification-policy map for one production workflow
- Early access to the GitHub App
- Direct founder support to configure one evidence path
- Founding-partner terms locked for 12 months
In return: one feedback call every two weeks, and permission to use anonymized product feedback.
Best for teams with 5–50 engineers · No credit card
Change surfaces
Evidence