AI Candidate Ranking System

Overview

This project ranks candidates for a Retrieval, Search, Recommendation, and Machine Learning Engineering role using a hybrid scoring system that combines:

• Feature Engineering
• Semantic Similarity
• Career Relevance Analysis
• Behavioral Signals
• Rule-Based Filtering

The system processes large candidate datasets and generates a ranked list of candidates based on overall suitability for the target role.

---

Problem Statement

Given:

• Candidate profiles
• Career history
• Skills
• Behavioral signals
• Job description

Rank candidates according to their relevance for a Retrieval/Search/Recommendation Engineering position.

---

Solution Architecture

Candidate Data ↓ Feature Extraction ↓ Semantic Matching ↓ Career Matching ↓ Score Aggregation ↓ Ranking ↓ Submission CSV

---

Features Used

Experience Score

Rewards candidates with:

• 5–9 years experience (highest score)
• 3–5 years experience
• 9–12 years experience

---

Skill Score

Matches candidate skills against:

• Retrieval
• Ranking
• Search
• Recommendation
• Embeddings
• Vector Databases
• LLM
• NDCG
• MRR
• Pinecone
• Qdrant
• Milvus
• Weaviate
• Elasticsearch
• OpenSearch

---

Career Score

Measures keyword relevance across the candidate's complete career history.

---

Title Score

Rewards titles such as:

• AI Engineer
• Machine Learning Engineer
• ML Engineer
• Applied Scientist
• Search Engineer
• Ranking Engineer
• Recommendation Systems Engineer
• Data Scientist

---

Title Relevance

Additional binary relevance signal based on AI/Search/ML-related titles.

---

Behavioral Score

Uses:

• Open To Work status
• Recruiter Response Rate
• GitHub Activity
• Profile Completeness
• Interview Completion Rate

---

Semantic Matching

The system uses:

Model: all-MiniLM-L6-v2

to compute semantic similarity between:

Job Description ↔ Candidate Profile

using cosine similarity.

---

Career Semantic Matching

Candidate career history is converted into embeddings and compared against a retrieval/search-focused query containing concepts such as:

• Search Systems
• Recommendation Systems
• Ranking
• Retrieval
• Semantic Search
• Embeddings
• Vector Databases
• LLM Ranking

---

Final Scoring Formula

Final Score =

0.20 × Career Score + 0.15 × Career Semantic Score + 0.20 × Behavior Score + 0.15 × Title Score + 0.10 × Skill Score + 0.10 × Experience Score + 0.10 × Semantic Score

---

Filtering Logic

Candidates with unrelated titles receive penalties.

if title_relevance == 0:
    score *= 0.5

Candidates with weak retrieval/search background receive additional penalties.

if career_score + career_semantic_score < 0.15:
    score *= 0.4

Performance Optimizations

Implemented optimizations:

• Cached JD Embedding
• Cached Career Query Embedding
• Batch Candidate Embeddings
• Batch Career Embeddings
• Single-pass Scoring

These optimizations significantly reduce runtime when processing 100,000 candidates.

---

Project Structure

resume-ranking-system/
│
├── data/
│   ├── candidates.jsonl
│   └── job_description.txt
│
├── src/
│   ├── candidate_text.py
│   ├── career_match.py
│   ├── export_results.py
│   ├── feature_engineering.py
│   ├── load_data.py
│   ├── model_loader.py
│   ├── ranker.py
│   ├── reasoning.py
│   ├── scorer.py
│   └── semantic_ranker.py
│
├── submission.csv
├── requirements.txt
├── README.md
└── main.py

Running

Install dependencies:

pip install -r requirements.txt

Run:

python main.py

Output:

submission.csv

containing:

• candidate_id
• rank
• score
• reasoning