FairTalent-Engine ⚖️⚡

An enterprise-grade, high-throughput Fair Machine Learning (Fair ML) auditing pipeline and psychometric integrity engine designed for AI-assisted talent acquisition.

Audits automated recruitment systems against the EEOC Four-Fifths Rule (29 CFR § 1607.4(D)), NYC Local Law 144 (AEDT Bias Audits), and EU AI Act Annex III (High-Risk AI in Employment), coupling Classical Test Theory construct validation ($\alpha \ge 0.70$) with automated Pareto threshold mitigation.

Designed and engineered by Fabio Torres (M.Sc. in Data Engineering & Cloud Infrastructure • 10+ Years Behavioral & Cognitive Psychology Leadership).

---

🧠 The Domain Moat: "Del Diván al Dato"

In automated recruitment and talent acquisition (HRTech), large-scale machine learning filters and ATS screening models routinely introduce latent discrimination against protected demographic groups (gender, age, ethnicity). Traditional algorithmic solutions fail because:

1. Software Engineers and Data Scientists know how to deploy XGBoost or vector databases, but do not understand psychometric construct validity, Classical Test Theory (CTT), item discrimination ($r_{bis}$), or federal employment law standards.
2. HR Practitioners and Industrial Psychologists understand Title VII and test administration, but cannot engineer distributed Polars pipelines, write vectorized async APIs, or automate post-processing mitigation algorithms.

FairTalent-Engine bridges this chasm:

graph TD
    subgraph BehavioralDomain ["Behavioral & Cognitive Science (10+ Years)"]
        B1["Classical Test Theory (Cronbach's α, SEM)"]
        B2["Item Discrimination & Construct Reliability"]
        B3["Title VII & EEOC Selection Protocol Standards"]
    end

    subgraph DataEngineering ["High-Performance Data Engineering"]
        D1["Pydantic v2 Contract Validation"]
        D2["Vectorized Polars Cohort Ingestion"]
        D3["Async High-Throughput FastAPI Microservices"]
    end

    subgraph FairnessDelivery ["Fairness ML & Regulatory Compliance"]
        F1["EEOC 80% Rule & Disparate Impact Ratio (DIR)"]
        F2["Pareto Optimal Post-Processing Threshold Calibration"]
        F3["Automated NYC LL144 & EU AI Act Audit Certificates"]
    end

    BehavioralDomain --> FairnessDelivery
    DataEngineering --> FairnessDelivery

    style BehavioralDomain fill:#0f172a,stroke:#34d399,stroke-width:1px,color:#f8fafc
    style DataEngineering fill:#0f172a,stroke:#38bdf8,stroke-width:1px,color:#f8fafc
    style FairnessDelivery fill:#1e293b,stroke:#a855f7,stroke-width:2px,color:#f8fafc

Loading

---

🏗️ System Architecture & Event Flow

flowchart LR
    subgraph Ingestion ["1. Applicant Ingestion"]
        AP["Applicant Records<br>• Demographics (Gender, Age, Race)<br>• Psychometric Test Items<br>• Candidate Merit Score"]
        VAL["Pydantic v2 Contracts<br><code>CandidateRecord</code>"]
        AP --> VAL
    end

    subgraph AuditCore ["2. Dual Audit Engine"]
        CTT["Psychometric Engine<br>• Cronbach's Alpha (α)<br>• Item-Total Correlation<br>• Standard Error (SEM)"]
        FAIR["Fair ML Auditor (Polars)<br>• Group Selection Rates<br>• Disparate Impact Ratio (DIR)<br>• Four-Fifths 80% Gate"]
        VAL --> CTT
        VAL --> FAIR
    end

    subgraph MitigationEngine ["3. Post-Processing Mitigation"]
        MIT["Threshold Calibrator<br>• Group Cutoff Tuning<br>• DIR ≥ 0.80 Resolution<br>• &gt;95% Merit Preservation"]
        FAIR -->|Adverse Impact Detected| MIT
    end

    subgraph Serving ["4. Serving & Compliance"]
        API["FastAPI Async REST API<br><code>/api/v1/audit/cohort</code><br><code>/api/v1/simulate/{scenario}</code>"]
        DASH["Interactive Off-White Web UI<br>1-Click Scenario Simulation & Telemetry"]
        CERT["Regulatory Audit Certificate<br>NYC LL144 & EU AI Act JSON"]
        
        FAIR --> API
        MIT --> API
        API --> DASH
        API --> CERT
    end

    style Ingestion fill:#0f172a,stroke:#38bdf8,stroke-width:1px,color:#f8fafc
    style AuditCore fill:#0f172a,stroke:#34d399,stroke-width:1px,color:#f8fafc
    style MitigationEngine fill:#0f172a,stroke:#f59e0b,stroke-width:1px,color:#f8fafc
    style Serving fill:#1e293b,stroke:#10b981,stroke-width:2px,color:#f8fafc

Loading

---

📐 Mathematical & Regulatory Foundations

1. EEOC Four-Fifths Rule (Disparate Impact Ratio)

Under 29 CFR § 1607.4(D), if the selection rate of a protected demographic group is less than 80% (4/5ths) of the highest selecting group, the selection tool possesses Adverse Impact:

$$\text{DIR} = \frac{\text{Selection Rate}_{\text{Minority Group}}}{\text{Selection Rate}_{\text{Reference Group}}} \ge 0.80$$

If $\text{DIR} < 0.80$, the system generates a Critical Legal Violation warning.

2. Psychometric Construct Reliability (Cronbach's Alpha)

Under Title VII employment litigation, an employer cannot defend a selection test unless it demonstrates proven construct validity:

$$\alpha = \frac{K}{K - 1} \left( 1 - \frac{\sum_{i=1}^K \sigma_i^2}{\sigma_X^2} \right)$$

Where $K$ is the item count, $\sigma_i^2$ is individual item variance, and $\sigma_X^2$ is aggregate test score variance. Baseline regulatory validity requires $\alpha \ge 0.70$.

3. Standard Error of Measurement ($SEM$)

Quantifies score uncertainty around decision cutoffs:

$$SEM = \sigma_X \sqrt{1 - \alpha}$$

---

🚀 Live Interactive Dashboard

The engine includes a zero-dependency, executive Off-White / Dark Mode Dashboard served directly at GET /:

• 1-Click Benchmark Cohort Simulation:
  • Biased Tech ATS: Simulates uncalibrated resume parsers penalizing female technical candidates ($\text{DIR} \approx 0.46$).
  • Executive Leadership Screening: Simulates automated tests penalizing candidates $\ge 40$ years old (ADEA violation).
  • Calibrated Pipeline: Demonstrates full compliance across all demographic categories.
• Interactive Cutoff Slider: Adjust selection cutoffs in real time ($50.0$ to $92.0$) and watch adverse impact ratios and selection rates update dynamically.
• Automated Fair-ML Mitigation Button: Automatically computes group-specific cutoffs to resolve the EEOC violation while retaining over $95%$ of original talent merit.
• Audit Certificate Exporter: Generates an official, structured compliance certificate aligned with NYC Local Law 144 and EU AI Act Annex III standards.

---

🔌 API Reference & Endpoints

Method	Endpoint	Description
GET	/	Serves the interactive executive web dashboard
GET	/health	System health check and regulatory framework status
GET	/api/v1/simulate/{scenario}	Simulates and audits a synthetic cohort (biased_tech_ats, age_penalized_exec, compliant_fair_pipeline)
POST	/api/v1/audit/cohort	Audits an uploaded batch of candidate records
POST	/api/v1/mitigate	Executes Pareto threshold calibration to eliminate adverse impact
GET	/api/v1/compliance/certificate	Generates a certified audit trail payload for legal compliance officers

Interactive OpenAPI documentation is available at /docs (Swagger) and /redoc.

---

🛠️ Quickstart Guide

1. Local Setup

# Clone repository
git clone https://github.com/neurodeveloper11/fairtalent-engine.git
cd fairtalent-engine

# Install dependencies
pip install -r requirements.txt

# Run test suite (100% pass)
python -m pytest -v tests/

# Launch development server
uvicorn src.main:app --host 0.0.0.0 --port 8000 --reload

Visit http://localhost:8000 for the dashboard and http://localhost:8000/docs for the interactive API explorer.

2. Docker Deployment

# Build and launch with Docker Compose
docker-compose up --build -d

# Check health status
curl http://localhost:8000/health

---

🧪 Testing & Validation Suite

The test suite covers psychometric accuracy, boundary edge cases, regulatory thresholds, and API handlers:

python -m pytest -v tests/

Test coverage includes:

• test_psychometrics.py: Verification of Cronbach's alpha, item discrimination, zero-variance handling, and Sten conversions.
• test_fairness_auditor.py: EEOC 80% rule validation, adverse impact detection, and regulatory synthesis.
• test_mitigation.py: Post-processing Pareto threshold tuning, merit score retention, and DIR resolution.
• test_generator.py: Deterministic synthetic cohort generation across standard hiring scenarios.
• test_api.py: FastAPI endpoints, error handling, contract validation, and certificate generation.

---

📄 License & Intellectual Property

Distributed under the MIT License. See LICENSE for details.

All applicant cohorts, candidate names, and simulation figures are 100% synthetic, generated for algorithmic demonstration and benchmark testing.

---

👤 Author & Lead Architect

Fabio Ignacio Torres Benítez
Data Engineer | Cognitive Scientist & Clinical/Organizational Psychologist | Full-Stack AI Developer
📍 Cali / Buenaventura, Colombia
🔗 LinkedIn | GitHub | Hugging Face | Google Play (NeuroGym Live)