CEREBRO: Contextual Evidence-Based Recruiter Optimization

(AI-powered candidate ranking system for intelligent recruitment) A hybrid multi-signal candidate ranking system that ranks candidates the way a great recruiter would — by understanding who fits the role, not just matching keywords.

Architecture Overview

100K Candidates
      │
      ▼
Phase 1: Fast Pre-Filter (heuristics, ~6s)
  ├─ Disqualify non-tech roles
  ├─ AI keyword presence scoring
  ├─ Basic YOE & activity filter
  └─ Top 5,000 shortlist
      │
      ▼
Phase 2: TF-IDF Semantic Similarity (~2s)
  └─ JD vs. candidate full text (title + summary + career + skills)
      │
      ▼
Phase 3: Multi-Signal Scoring (~2s per 5K)
  ├─ Hard Skill Coverage (6 groups: embeddings, vector DBs, eval, LLM, MLOps, Python)
  ├─ Career Quality (product company vs consulting, progression, tenure)
  ├─ Experience Fit (sweet spot: 6–8yr, JD range 5–9yr)
  ├─ Behavioral Signals (activity, response rate, notice period, GitHub)
  ├─ Location Fit (Pune/Noida > other JD cities > India > elsewhere)
  └─ Anti-Pattern Detection (honeypots, keyword stuffers, consulting-only, CV-only)
      │
      ▼
Weighted Ensemble → Top 100 → submission.csv

Key Design Decisions

Why NOT pure keyword matching

The JD explicitly warns: "The right answer is not finding candidates whose skills section contains the most AI keywords." CEREBRO reads career descriptions, not just skill lists. A candidate who built a recommendation system at a product company without listing "RAG" gets ranked higher than a title-match with hollow skill tags.

Anti-pattern detection (honeypot safety)

CEREBRO detects and penalizes:

• Honeypots: Impossible timelines (claimed 8yr exp at 3yr-old company), expert skills with 0 months usage
• Keyword stuffers: Many AI skill tags but no matching technical depth in career descriptions
• Consulting-only: Entire career at TCS/Infosys/etc — explicitly flagged in JD
• Pure researchers: Heavy academic signals with no production deployment evidence
• Title chasers: Short tenures (< 12mo) across last 3 roles
• CV/NLP mismatch: Computer vision specialists without NLP/IR background

Behavioral signals as availability multiplier

A perfect-on-paper candidate who hasn't logged in for 6 months is, for hiring purposes, not actually available. CEREBRO weights:

• Recency of last login (25%)
• Open-to-work flag (20%)
• Recruiter response rate (20%)
• Notice period (15%) — JD says sub-30 days preferred
• Interview completion rate (10%)
• GitHub activity (10%)

Two-phase architecture for <5 min CPU constraint

Full TF-IDF on 100K docs is slow. CEREBRO uses a fast heuristic pre-filter (pure Python, no ML) to narrow to 5K candidates, then applies the heavy scoring only on the shortlist. Result: ~17 seconds on a standard CPU.

Quick Start

Requirements

pip install numpy pandas scikit-learn scipy tqdm

Produce submission CSV

python rank.py --candidates ./candidates.jsonl --out ./submission.csv

Options

--candidates  Path to candidates JSONL file (default: candidates.jsonl)
--out         Output CSV path (default: submission.csv)
--top_n       Number of top candidates (default: 100)

Expected runtime

Phase	Time
Load 100K candidates	~7s
Pre-filter (heuristics)	~6s
TF-IDF on 5K shortlist	~2s
Full scoring	~2s
Total	~17s

Scoring Weights

Component	Weight	Rationale
TF-IDF Semantic Match	22%	Captures JD–profile alignment beyond keywords
Hard Skills Coverage	20%	Must-haves: embeddings, vector DBs, eval, Python
Career Quality	18%	Product company > consulting, progression, tenure
Behavioral Signals	15%	Availability, engagement, responsiveness
Anti-Pattern Penalty	9%	Inverted: deducts for honeypots/mismatches
Experience Fit	10%	6–8yr sweet spot, 5–9yr acceptable range
Location Fit	6%	Pune/Noida > other India cities > willing to relocate

File Structure

cerebro/
├── rank.py                    # Main ranker (single-command execution)
├── README.md                  # This file
├── requirements.txt           # Python dependencies
├── submission_metadata.yaml   # Hackathon metadata
└── submission.csv             # Generated output (top 100 ranked candidates)

Compute Constraints Compliance

• ✅ No hosted LLM APIs called during ranking
• ✅ CPU-only (no GPU required)
• ✅ <5 minutes runtime (~17 seconds actual)
• ✅ <16GB RAM (uses ~1–2GB peak for TF-IDF matrix)
• ✅ Single command produces submission CSV