AI Resume Screening and Job Fit Prediction System v2

This project predicts the best-fit job role from resume text and compares a resume with a job description. The upgraded pipeline uses a repository-local Sentence Transformer, XGBoost, spaCy EntityRuler skill extraction, ONNX Runtime, and FAISS instead of TF-IDF and hardcoded keyword matching.

What Changed

• download_model.py saves all-MiniLM-L6-v2 locally to models/all-MiniLM-L6-v2.
• Resume text is converted into dense vectors with the local Sentence Transformer model.
• Job-role classification is trained with XGBClassifier on those dense embeddings.
• Rare job roles are handled with compute_sample_weight(class_weight="balanced") during training.
• Skill extraction uses spaCy EntityRuler patterns loaded from data/skills_patterns.json.
• JD semantic matching embeds text with the ONNX-preferred local model and searches job-description vectors with FAISS IndexFlatIP.

Dataset

Training data is expected at:

data/resume_data.csv

The target column is:

job_position_name

The model input is a leakage-free profile_text field built from resume-side columns when present:

skills
career_objective
responsibilities
positions
degree_names
major_field_of_studies
educational_institution_name

Project Structure

resume-screening-ml-v2/
|-- app/
|   |-- streamlit_app.py
|-- data/
|   |-- resume_data.csv
|   |-- skills_patterns.json
|-- models/
|   |-- all-MiniLM-L6-v2/
|   |-- sentence_transformer_onnx/
|   |-- job_role_classifier.pkl
|   |-- job_role_label_encoder.pkl
|   |-- job_role_embedding_config.json
|   |-- job_role_model_report.json
|-- reports/
|-- src/
|   |-- embedding_model.py
|   |-- export_onnx_model.py
|   |-- jd_matcher.py
|   |-- predict.py
|   |-- preprocessing.py
|   |-- skill_extractor.py
|   |-- train_model.py
|   |-- utils.py
|-- download_model.py
|-- requirements.txt

Install And Prepare Models

pip install -r requirements.txt
python download_model.py
python src/export_onnx_model.py

download_model.py skips work if models/all-MiniLM-L6-v2 already exists. The ONNX export is used by JD matching for faster CPU inference.

Training

python src/train_model.py

Training now does this:

1. Loads and cleans data/resume_data.csv.
2. Builds leakage-free resume profile_text.
3. Removes classes with fewer than 2 samples so stratified splitting is valid.
4. Encodes resume text with models/all-MiniLM-L6-v2.
5. Computes balanced sample weights for imbalanced job-role classes.
6. Trains XGBoost with multi:softprob.
7. Saves the classifier, label encoder, embedding config, and evaluation reports.

models/job_role_vectorizer.pkl has been removed and is no longer saved or loaded.

Running The App

streamlit run app/streamlit_app.py

The app shows the predicted job role, confidence, top 3 roles, semantic JD similarity, skill match percentage, matched skills, missing skills, and final recommendation.

JD Matching

JD matching embeds the resume and job description with the local ONNX Sentence Transformer, normalizes the vectors, creates a FAISS IndexFlatIP index for job descriptions, and searches that index for the highest semantic match.

The final score is:

final_jd_score = 0.7 * semantic_similarity_score + 0.3 * skill_match_percentage

Skill overlap is based on spaCy EntityRuler entities loaded from data/skills_patterns.json.

Smoke Tests

python -c "from src.predict import predict_job_role; print(predict_job_role('Python React Docker AWS machine learning engineer'))"

python -c "from src.jd_matcher import calculate_jd_match_score; print(calculate_jd_match_score('Python Docker AWS ML', 'Need Python, AWS, Docker'))"

python -c "from src.skill_extractor import extract_skills; print(extract_skills('Built APIs with Python, FastAPI, Docker, Kubernetes and AWS'))"

Notes

• The trained local XGBoost model currently reports accuracy 0.8591 and weighted F1 0.8627 on the validation split.
• Model quality depends on the coverage and cleanliness of data/resume_data.csv.
• xgboost, spacy, faiss-cpu, and onnxruntime are required runtime dependencies.