HireNFire

HireNFire -> Bias Aware Hiring Screener

OpenEnv Environment where an AI agent screens resumes, conducts structured interviews, and makes hiring decisions graded on both quality of hire AND demographic fairness.

---

Motivation

Hiring is one of the most consequential decisions organizations make. An AI agent that can assist in screening must navigate a fundamental tension:

reward = α · quality + (1 − α) · fairness

• Optimizing purely for the "best" candidate often produces demographic skew
• Optimizing for demographic parity alone produces weaker hires
• The agent must learn to balance both objectives simultaneously

This makes the environment scientifically interesting and practically relevant.

---

Architecture

┌─────────────────────────────────────────┐
│            Agent (LLM via OpenAI)       │
└──────────────────┬──────────────────────┘
                   │ action
                   ▼
┌─────────────────────────────────────────┐
│         step(action) → obs, reward,     │──→ Reward Grader
│                        done, info       │
└──────────────────┬──────────────────────┘
                   ▼
┌─────────────────────────────────────────┐
│                 State                    │
│  ┌──────────┐ ┌───────────┐ ┌────────┐ │
│  │Candidate │ │ Decision  │ │ Task   │ │
│  │  Pool    │ │   Log     │ │ Config │ │
│  └──────────┘ └───────────┘ └────────┘ │
└──────────────────┬──────────────────────┘
                   ▼
     ┌─────────┐ ┌──────────┐ ┌───────────┐
     │Quality  │ │Fairness  │ │  Final    │
     │Score    │ │Score     │ │  Reward   │
     │P/R      │ │Parity   │ │α·Q+(1-α)·F│
     └─────────┘ └──────────┘ └───────────┘

---

Observation Space

Each observation contains:

Candidate Fields

---

Action Space

---

Tasks

Task 1: Simple Ranking (Easy)

• Role: Software Engineer
• Candidates: 10 with clear skill gaps
• α = 0.7 (quality weighted)
• Goal: Rank and hire the top candidates by explicit criteria
• Expected Score: 0.7–0.9

Task 2: High-Potential Screening (Medium)

• Role: Data Scientist
• Candidates: 25, some underqualified but high potential
• α = 0.5 (balanced)
• Goal: Balance strict criteria with potential; maintain fairness
• Expected Score: 0.5–0.7

Task 3: Fair Hiring at Scale (Hard)

• Role: Senior ML Engineer
• Candidates: 50 with demographic skew and tight margins
• α = 0.3 (fairness weighted)
• Goal: Maintain demographic parity while hiring quality candidates
• Expected Score: 0.3–0.6

---

Setup & Usage

Prerequisites

• Python 3.11+
• OpenAI API key (for baseline inference)

Installation

# Clone or download
cd HireNFire

# Install dependencies
pip install -r requirements.txt

# Set API key (for baseline inference only)
export OPENAI_API_KEY="your-key-here"

Run the Notebook

jupyter notebook HireNFire.ipynb

Docker

docker build -t hirenfire .
docker run -p 8888:8888 hirenfire

Hugging Face Spaces

Tagged with openenv. Deploy as a Docker Space:

# Push to HF Spaces
# Tag: openenv

---

Baseline Scores

Heuristic agent (rank by computed score → hire top-K):

Scores are approximate and will vary slightly due to random candidate generation.

---

Why This Environment is Interesting

1. Genuine tradeoff: Quality and fairness are partially conflicting objectives
2. Real-world relevance: Hiring bias is a well-studied, impactful problem
3. Rich action space: Multiple action types allow diverse strategies
4. Partial progress signals: Intermediate rewards for good interview decisions
5. Configurable tension: α parameter controls the quality–fairness tradeoff