Issue with my Thesis Hey everyone, I’m currently working…
0 reactions
Issue with my Thesis
Hey everyone,
I’m currently working on my bachelor thesis in collaboration with a company and ran into a conceptual issue that I’d like some input on.
The topic is about using LLMs for code reviews (analyzing code changes (diffs), relating them to a ticket or user story, and generating meaningful feedback beyond classic static analysis).
Here’s the issue:
* The company requires a fully local setup (no external APIs like OpenAI/Anthropic) due to cost and policy constraints.
* My professor is very sceptical about this approach. His main concern is that local models won’t be capable enough (especially when it comes to handling larger contexts (ticket + diff + relevant codebase parts)) and actually reasoning about whether requirements are correctly implemented.
His argument is basically:
If the system can’t go beyond shallow analysis, it risks becoming “static analysis + some NLP,” which wouldn’t be sufficient for a bachelor thesis.
So I'm kinda stuck here.
Do you think this setup is fundamentally too limited, or is there still a viable direction here?
I’m not looking for implementation help, but more for:
* conceptual approaches that could make this non-trivial
* ways to structure the problem so local models are sufficient
* or whether his concern is realistically justified
Curious if anyone here has worked on LLMs in constrained environments or has thoughts on whether this is a dead end or not.
**TL;DR:**
Bachelor thesis on LLM-based code reviews. Company requires local models only, professor doubts they’re strong enough → risk of trivial outcome. Looking for perspectives on whether this can still be a solid research topic.
#technology
source
Hey everyone,
I’m currently working on my bachelor thesis in collaboration with a company and ran into a conceptual issue that I’d like some input on.
The topic is about using LLMs for code reviews (analyzing code changes (diffs), relating them to a ticket or user story, and generating meaningful feedback beyond classic static analysis).
Here’s the issue:
* The company requires a fully local setup (no external APIs like OpenAI/Anthropic) due to cost and policy constraints.
* My professor is very sceptical about this approach. His main concern is that local models won’t be capable enough (especially when it comes to handling larger contexts (ticket + diff + relevant codebase parts)) and actually reasoning about whether requirements are correctly implemented.
His argument is basically:
If the system can’t go beyond shallow analysis, it risks becoming “static analysis + some NLP,” which wouldn’t be sufficient for a bachelor thesis.
So I'm kinda stuck here.
Do you think this setup is fundamentally too limited, or is there still a viable direction here?
I’m not looking for implementation help, but more for:
* conceptual approaches that could make this non-trivial
* ways to structure the problem so local models are sufficient
* or whether his concern is realistically justified
Curious if anyone here has worked on LLMs in constrained environments or has thoughts on whether this is a dead end or not.
**TL;DR:**
Bachelor thesis on LLM-based code reviews. Company requires local models only, professor doubts they’re strong enough → risk of trivial outcome. Looking for perspectives on whether this can still be a solid research topic.
#technology
source
| 3,000 mlx | |
| 3 | |
| 0 |