Few suggestions for improving the Resume Matcher

[wakil69]

Suggestions for Improving the Resume Matcher Project

Hi!

First of all, amazing work on this project.
The job market is tough, and tools like this can truly help people improve their chances.

After exploring the repository (and drawing from my experience building [Play2Path](https://play2path.com): a resume generator and job board), I’d like to share a few improvement ideas and would be happy to contribute.

---

1. Dependency version issue

There seems to be a problem with the current API version.
I was able to fix it locally by updating the dependency version in pyproject.toml to:

fastapi==0.119.0

---

2. Local LLM handling

The local LLM setup doesn’t seem to function properly, while API-based models (like gpt-4o) work as expected.
Adding clearer support for local inference backends (e.g., Ollama, LM Studio, or open-source models) would improve flexibility for users without API access.

---

3. Privacy improvements

Before sending resume data to external LLM providers, it would be safer to remove personal information such as phone numbers and email addresses.
These fields aren’t relevant for content improvement and should stay local to enhance privacy and compliance.

---

4. Multilingual support

Currently, the project supports only English.
Adding multi-language support (starting with French, Spanish, and German) could significantly expand usability and adoption (I am French).

---

5. Resume PDF export

It would be great to allow exporting the enhanced resume as a PDF, either from the dashboard or via an API endpoint.

---

6. Structured JSON extraction

Right now, job and resume parsing are done in a single LLM call, which can cause parsing errors or incomplete data.
It might be better to split the LLM calls by section, for example:

• Experience
• Education
• Skills / Miscellaneous

This more granular approach would improve reliability and reduce schema validation failures.

---

7. Similarity scoring enhancement

The current similarity computation is between:

• The job’s extracted keywords embedding, and
• The entire raw resume embedding

This often yields low scores, since the resume text includes a lot of irrelevant content.

Possible improvements:

1. Use the LLM to directly assess semantic similarity between job keywords and resume.
2. Compute a weighted average of cosine similarity across sections (skills, experience, education), increasing the score if job keywords appear explicitly.

---

8. Personalized improvement suggestions

A great next step would be to generate personalized career recommendations, such as:

• Suggested courses or certifications
• Missing skills to learn
• Personal project ideas relevant to the job field

This could make the tool a career guidance assistant, not just a resume matcher.

---

I’d be happy to help with these improvements, especially the embedding and LLM handling parts.

Thanks again for your incredible work, this project has huge potential to make job applications smarter and fairer.