Re: request for help: LLM-based quality assurance

Hi,

On 6/5/26 17:03, Pavel Cahyna wrote:
> Here are the findings:
...


The results look great!

Do you mind providing details about which provider/model and which 
prompt(s) you were using?

In general, I am very interested in using LLMs to improve code quality.
So far I had great success in reviewing MRs / PRs with different Gemini 
and Claude models.
But it quickly gets expensive, so I switched using Qwen3.6-35B on my 
local development machine (AMD M780), used with the `pi` agent.
It's slow and quality is lower (but not too bad) in comparison to 
high-end paid models.
It found two buffer overflows in the good ol' whois utility.

On Gitlab, gnulib has the Ultimate tier, which allows us to use 
instances with a couple of CPU cores and RAM - maybe enough to run the 
the LLM on CPU only.
Just mentioning this as an optional alternative to external models.

Regards, Tim

-----BEGIN PGP SIGNATURE-----
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=RAkJ
-----END PGP SIGNATURE-----