Redrob Hackathon — Ranker

Team: ksploitx
Author: Khushneet Singh

Quick Start

pip install -r requirements.txt   # no extra deps beyond stdlib
python rank.py --candidates ./candidates.jsonl --out ./submission.csv
python validate_submission.py submission.csv

Runtime: ~50s on CPU (100K candidates). No GPU, no network calls.

Architecture

Rule-based multi-component scorer — no LLM API calls, no GPU required.

Scoring Components (weighted)

Component	Weight	What it measures
Skills match	30%	Proficiency × trust (endorsements + duration) across must-have and nice-to-have skills
Career fit	28%	Product company history, AI/ML titles, retrieval/ranking evidence in job descriptions
Experience	15%	Years of experience (5-9 ideal per JD)
Behavioral signals	12%	Recency, open_to_work, recruiter response rate, notice period
Location	10%	Noida/Pune > major Indian cities > India > willing to relocate
Education	5%	Institution tier + CS/ML field bonus

Anti-signals (multiplicative penalties)

• Entire career at consulting firms (TCS/Infosys/Wipro etc.): ×0.5
• CV/speech/robotics-only profile with no retrieval exposure: ×0.6
• Non-technical current title (marketing, sales, etc.): ×0.4
• International location, not willing to relocate: ×0.35
• Excessive job-hopping (3+ stints <18 months): ×0.85

Honeypot Detection

Profiles with inconsistencies (skill proficiency vs duration mismatch, expert skills with 0 endorsements, impossible career timelines, non-tech titles with AI keyword stuffing) are scored near 0 and excluded from top 100.

Key Design Decisions

1. Career descriptions over keyword lists — skills listed are checked against actual job descriptions for trust/endorsement backing. A "expert" skill with 0 endorsements and 0 duration months is penalized.

2. Product company preference — the JD explicitly wants non-consulting backgrounds. Company names and industries are checked against a consulting firm list.

3. Behavioral signals as multiplier — a perfect-on-paper candidate who hasn't logged in for 6 months or has 5% response rate ranks lower than a slightly-less-perfect but engaged candidate.

4. No API calls — fully offline, CPU-only, runs in <60 seconds.