Yuvaakhil18/redrob-ranker

🎯 RedRob Intelligent Candidate Ranking System

    

India Runs Data and AI Challenge | Challenge 1 | Hack2Skill × Redrob An enterprise-grade, offline-online hybrid semantic and rule-based candidate discovery and ranking system. Designed to search, normalize, score, and justify candidate recommendations from a pool of 100,000+ applicants under 20 seconds.

Live Space Link: https://huggingface.co/spaces/Yuvaakhil18/redrob-ranker GitHub Repository: https://github.com/Yuvaakhil18/Redrob-Intelligent-Candidate-Ranking-System

🏗️ Architecture

The system splits computation into a Phase 1 Offline Precomputation (to generate high-quality text embeddings) and a Phase 2 Online Pipeline (to rank 100K candidates under 20 seconds within a restricted timed CPU window).

flowchart TD
  subgraph Offline [Phase 1: Offline Precomputation]
    A[candidates.jsonl] --> B(all-MiniLM-L6-v2 Embedder)
    B --> C[candidate_embeddings.npy]
  end
  subgraph Online [Phase 2: Online Ranking Pipeline <20s]
    JD[Job Description] --> D(JD Embedder)
    D --> E[jd_embedding.npy]
    C --> F(Memory-Mapped Cosine Similarity)
    E --> F
    Raw[Raw Profiles Stream] --> G(Normalization & 5-Dim Scoring)
    F --> G
    G --> H{Honeypot Filter}
    H -- Fraud --> I[Penalty Cap 0.05]
    H -- Normal --> J[Behavioral Multiplier]
    I --> K(Deterministic Sorting)
    J --> K
    K --> L[Rule-Based Reasoner]
    L --> Out[submission.csv]
  end

👥 Team Information: Jutsu Engineers

• —Team Name: Jutsu Engineers
• —Primary Contact: C.Yuvaakhil (yuvaakhil2318@gmail.com)
• —Team Members:
• —C.Yuvaakhil (Leader)
• —Bathula Laxman Karthik
• —Abburi Likitha Nalini
• —SATYANARAYANA SANKA

🎯 How this maps to the Evaluation Rubric

📈 Ranking & Scoring Logic

The final candidate rank is determined by combining semantic and profile-based attributes:

$$\text{Final Score} = \text{Clamp}\left(\sum{d=1}^{5} (\text{Score}d \times \text{Weight}_d), 0.0, 1.0\right) \times \text{Behavioral Multiplier} \quad [\text{Capped at 1.2}]$$

1. Dimension 1: Role Match (Weight: 30%)

• —Semantic Vector Cosine Similarity (50%): Measures semantic relevance between candidate profile text and the Job Description.
• —Company Type Check (20%): Flags candidates with experience exclusively in service/IT consulting companies (penalized to a 0.4 score factor, product/startup backgrounds get a 1.0).
• —Keyword Matching (30%): Checks for critical technical terms (retrieval, embedding, embeddings, ranking, vector, rag) in career history text.

2. Dimension 2: Skills Validation (Weight: 20%)

• —Checks for mandatory skills (python, embeddings, vector database, retrieval) and optional skills (llm fine-tuning, evaluation frameworks, ranking).
• —Individual skill scores are mapped using proficiency level, duration, and endorsement counts.
• —Required Skills Threshold: Candidates scoring $<0.4$ on required skills are filtered out (receive 0.0).

3. Dimension 3: Experience Level Fit (Weight: 15%)

• —Matches total years of experience against the optimal range of 6–8 years (Score: 1.0).
• —Other ranges degrade gracefully (e.g., 5-6 / 8-9 years: 0.85; 4-5 / 9-12 years: 0.70; $>12$ years: 0.40; $<4$ years: 0.30).
• —Startup/Scaleup Bonus: $+10\%$ bonus added if startup/scaleup environments are in their history.

4. Dimension 4: Location & Availability (Weight: 15%)

• —Location Scoring: Prefers candidates in Indian tech hubs (Pune, Noida, Delhi, Hyderabad, Bangalore, Mumbai) who are willing to relocate.
• —Notice Period Scaling: Shorter notice periods receive higher multipliers ($\le30$ days: 1.0; $\le60$ days: 0.85; $\le90$ days: 0.60; $>90$ days: 0.30).

5. Dimension 5: Education (Weight: 5%)

• —Mapped by degree level and computer science (CS) specialization:
• —PhD: 0.9.
• —Masters: 1.0 (CS) / 0.8 (Non-CS).
• —Bachelors: 1.0 (Tier-1/2 CS) / 0.85 (Tier-3 CS) / 0.5 (Non-CS).
• —Others/Self-taught: 0.6 default.

6. Dimension 6: Behavioral Multiplier (Scale: 0.4x to 1.2x)

Platform engagement signals dynamically adjust the base score:

• —Actively Open to work: $+0.1$
• —Platform active within last 30 days: $+0.15$
• —Recruiter response rate $\ge70\%$: $+0.1$ (penalized $-0.15$ if $<30\%$)
• —Interview completion rate $\ge80\%$: $+0.05$ (penalized $-0.1$ if $<50\%$)
• —Active GitHub score (>30): $+0.05$
• —Verified Email & Phone: $+0.05$
• —Offer acceptance rate $\ge80\%$: $+0.05$ (penalized $-0.05$ if $<30\%$)

🛡️ Honeypot & Fraud Detection

To safeguard ranking integrity, candidate profiles are scanned against six fraudulent heuristic patterns:

1. 1.Zero-Duration Experts: Profiles claiming "expert" proficiency in a skill with 0 months of duration.
2. 2.Zero-Validation Experts: Advanced/Expert skills listed with 0 endorsements.
3. 3.Experience Mismatches: Stated total years of experience differs from reconstructed employment history timelines by $>10\%$ or $>1$ year.
4. 4.Skill Stuffing: 8 or more expert-level skills claimed with 0 endorsements.
5. 5.Empty History Claims: Stating positive years of experience but having zero career history entries.
6. 6.Incomplete Profiles: Profile completeness scores below 40%.

Action Taken: If a candidate triggers critical mismatch patterns or accumulates $\ge3$ minor flags, they are classified as a honeypot. Their final score is immediately capped at a flat 0.05, keeping them at the bottom of the list and keeping your top 100 recommendation list completely clean.

🚀 Quick Start

1. Installation

# Set up virtual environment
python -m venv venv
source venv/Scripts/activate  # On Windows: venv\Scripts\activate

# Install requirements
pip install -r requirements.txt

2. Run Offline Precomputation

python precompute.py --candidates candidates.jsonl.gz --output-dir embeddings_output

3. Run Real-Time Online Ranking

python rank.py --candidates candidates.jsonl.gz --embeddings embeddings_output --out submission.csv

4. Running Tests & Validation

To verify pipeline components and execution:

python tests/test_all.py

To validate the CSV structure against hackathon rules:

python validate_submission.py submission.csv