AI-powered hiring decisions with built-in bias detection, candidate scoring, and CEO-ready PDF reports.
Built in 6 days using Python, FastAPI, Claude AI (Anthropic), and Streamlit.
Most companies are adopting AI in hiring faster than they are auditing it for fairness.
67% of companies now use AI in hiring. Only 26% of candidates trust it.
This platform bridges that gap — it makes AI hiring decisions faster AND documents them for bias, compliance, and audit purposes.
| Module | What It Does |
|---|---|
| 🔍 Analyse Problem | Paste any HR problem in plain English → AI extracts urgency, constraints, hidden risks |
| 👥 Score Candidates | Role-based weighted scoring (technical / executive / operational) with confidence bands |
| ✅ Create Decision | AI-powered APPROVE / REJECT with reasoning and next step recommendation |
| 🚨 Detect Bias | Scans job descriptions and interview notes for 8 bias types with legal risk and suggested rewrites |
| 📄 PDF Report | One-click 4-page CEO-ready report with full audit trail |
Feed it a real job description:
"We need a young energetic rockstar developer who fits our culture. Recent CS grad from top university preferred."
Result in 3 seconds:
- 🔴 Age Discrimination (HIGH) — trigger: "young" — violates ADEA
- 🟠 Educational Elitism (MEDIUM) — trigger: "top university" — creates disparate impact
- 🟠 Culture Fit Vagueness (MEDIUM) — known legal proxy for racial/gender discrimination
Each flag includes: exact trigger phrase, legal explanation, compliance risk, and suggested rewrite.
Backend: Python 3.13 + FastAPI
AI Engine: Claude Haiku (Anthropic API)
Frontend: Streamlit
PDF: ReportLab
Server: Uvicorn
git clone https://github.com/YOUR_USERNAME/ai-hr-decision-platform.git
cd ai-hr-decision-platformpip install fastapi uvicorn anthropic python-dotenv streamlit reportlabCreate a file called .env in the project folder:
ANTHROPIC_API_KEY=your_api_key_here
Get your free API key at: console.anthropic.com
uvicorn main:app --reloadWait for: Application startup complete
streamlit run app.pyhttp://localhost:8501
All endpoints available at http://localhost:8000/docs (Swagger UI)
| Method | Endpoint | Description |
|---|---|---|
| GET | /health |
Server status check |
| POST | /analyse-problem |
Extract structured data from HR problems |
| POST | /score-candidates |
Role-weighted candidate ranking |
| POST | /create-decision |
AI hiring decision with reasoning |
| POST | /detect-bias |
Scan content for bias patterns |
| POST | /audit-decisions |
Systemic bias analysis across decision batches |
| POST | /bias-score-candidates |
Scoring with integrated bias check |
| POST | /generate-report |
Generate CEO-ready PDF report |
- 67% of companies now use AI in hiring (up from 26% in 2024)
- Only 26% of candidates trust AI to evaluate them fairly (Gartner 2026)
- New York City legally requires annual bias audits before using any AI hiring tool
- California extended anti-discrimination laws to cover AI hiring tools in 2025
- EU AI Act classifies hiring AI as high-risk — requiring documentation and human oversight
Your PDF audit trail is the documentation these laws require.
ai-hr-decision-platform/
│
├── main.py # FastAPI backend — all API endpoints
├── app.py # Streamlit frontend — visual UI
├── report.py # ReportLab PDF generator
├── .env # Your API key (NOT uploaded to GitHub)
├── .gitignore # Prevents .env from being uploaded
└── README.md # This file
Your .env file containing your Anthropic API key is excluded from this repository via .gitignore.
Never commit API keys to GitHub. If you accidentally do, rotate your key immediately at console.anthropic.com.
- User authentication and multi-tenancy
- PostgreSQL database for decision history
- Interview question generator
- Offer letter drafting module
- Knowledge layer — learns from company's own hiring history
- Workday / Greenhouse / Lever integrations
- SOC 2 compliance
Built by a system thinker exploring the intersection of AI, HR compliance, and enterprise decision-making.
Connect on LinkedIn: your-linkedin-url
Follow on X: your-x-handle
MIT License — free to use, modify, and build on.
Built in 6 days. Under 500 lines of code. Powered by Claude AI.