Personalized Resume & Cover Letter Generator

IDE

Download Pycharm: https://www.jetbrains.com/pycharm/download/?section=windows or https://www.jetbrains.com/pycharm/download/?section=mac

How to use Pycharm

1. Open pycharm and create new project

2. Create Virtual env with python3.12

Setup

1. Open Pycharm and create virtual environment using python3.12
2. pip install -r requirements.txt
3. Get OpenAI API key
4. Create .env file inside project and paste OPENAI_API_KEY="sk-..." in your .env
5. streamlit run app/app.py
6. We will also perform fine-tuning of model today. Set GEN_MODEL=ft:gpt-4o-mini-2024-07-18:personal:resume-cover-ft:C3HhrPnR post finetuning job in .env
7. Post finetuning is done, we will run streamlit run scripts/ab_test_UI.py : https://platform.openai.com/docs/guides/supervised-fine-tuning

Features:

1. Upload/Paste Job description and Resume Deatils

2. Provide Few-shot Examples ( Optional ): https://learnprompting.org/docs/basics/few_shot?srsltid=AfmBOopYaZb7zlugXPrUWgLXxuUTpOLuDn8zEJtmXHpYZqdu4o_OD_O9. https://platform.openai.com/docs/guides/prompt-engineering

3. Generate tailored bullet pointers for resume

4. Generate tailored cover letter

5. Enhace the ouptput with few-shot examples

6. Fine-tuning of GPT-4o-min

7. Compare performance and store the results

8. Key metrics Definition:

   a. keyword_coverage Fraction of important JD terms that show up in the model’s output (0–1 scale). We extract likely keywords (TitleCase, long words, common skills) and check presence. Higher = better JD alignment.

   b. quantify_score Measures how “quantified” the writing is by counting numbers, %, and impact verbs per line. Higher means more measurable outcomes (great for resume bullets).

   c. length_ok (cover letters only) Boolean check that the letter lands in a sensible range (default 120–200 words). Helps keep letters concise yet substantial.

   d. composite_score Single yardstick combining the above: for bullets → ~60% keyword coverage + 40% quantification; for cover letters → ~50% keyword + 40% quant + small bonus if length_ok. Capped at 1.0 for easy comparison.

Data preparation for fine-tuning

Steps:

1. python scripts/prep_dataset.py This step will create finetune.jsonl
2. python scripts/run_finetune.py This will execute a finetuning job and will create data/tuned_model.txt