Launching today
Abloh
Verify AI-written code and tests on every pull request
3 followers
Verify AI-written code and tests on every pull request
3 followers
Push your branch. Open the pull request. Abloh runs automatically. On every changed line, it runs a combination of mutation testing and invariant testing with AI triage to verify your tests and allow you to actually find bugs. You get a per-line verdict on which changes are protected and where the gaps are: untested code, tests that pass whatever the code does, edge cases nobody wrote a test for. Merge AI-written code on evidence instead of assumptions.
Hey Product Hunt! 👋 I'm Logan, co-founder of Abloh.
In my Co-founder's last role as a SWE, the company laid off most of its QA team and testing fell to developers who'd never specialized in it. That quarter became the worst for production bugs in the company's history. Everyone was shipping faster than they could verify anything.
AI made that gap wider. Across 470 pull requests we studied, AI-written changes had twice the rate of missing error handling. Coding agents only test the paths they implemented: if the agent never handled an empty response or a failed request, there's no test for it either. Existing coverage tools are useless in this case.
Test quality can only be measured by making the code wrong. That's mutation testing, and it's been the known answer for decades. It stayed unused because it was too noisy (Google measured 85% of mutants as unproductive false positives) and because it can't plant a fault in code that was never written.
So we fine-tuned a model to fix both:
🔇 60% of the noise suppressed, so surviving mutants are worth reading
🎯 Across 2,500 production bugs, it flags 80% before they ship
📊 20% above Opus 5 and 10% ahead of GPT-5.6 Sol high
✅ A per-line verdict on every PR showing which changes are genuinely protected
Abloh isn't a coverage tool or another linter. It's verification: it proves your tests would fail if the code broke, so AI-written code gets merged on evidence instead of assumptions.
I'd love your feedback, especially from engineers reviewing agent-written PRs. How are you deciding what's safe to merge right now? Happy to answer anything! 🚀🚀