How I built a multi-LLM consensus tool to stop AI from hallucinating in my code
AI models disagree on your code. NexaVerify shows what they agree on.
I ran the same code through multiple AI models…
and they gave completely different answers.
One said “no issues”.
Another flagged a security problem.
That’s when I realized something:
AI code review isn’t reliable if you only use one model.
So I built a small tool that runs code through multiple AIs
and highlights what they actually agree on.
It’s still early, but it already caught things a single AI missed.
Curious — has anyone else noticed this?
Do you trust one AI for code review?
→ Free trial (3 scans/day, no credit card):
nexadiag.gumroad.com/l/nexaverifyfree
Would love feedback from anyone who tries it.
Update: I got the first version live on Gumroad.
Free tier is available (3 scans/day, no credit card needed,
Groq works for free).
The consensus approach surprised me — on my own codebase,
running 4 models caught 2 critical bugs that GPT alone missed.
Still 0 sales but the product works.
Would genuinely love a tester if anyone wants to try it free.
→ nexadiag.gumroad.com/l/nexaverifyfree
I'm currently thinking about the next step: is it better to add more 'lightweight' models for speed, or stick to the heavy hitters?
More importantly, I’m refining the consensus criteria itself. Most tools on the market right now are honestly all over the place and often give unreliable results. Getting the verification engine's logic right is my top priority.
If you have specific preferences on the models you use—or if you've been frustrated by current tools telling you nonsense—I’d love to hear your thoughts!
This comment was deleted 5 months ago