import gradio as gr
import pandas as pd
import json

# Load KPIs
with open("kpis.json", "r") as f:
    kpis = json.load(f)

# Load tables
audit_df = pd.read_csv("hiring_audit_results.csv")
subgroup_df = pd.read_csv("subgroup_score_analysis.csv")

kpi_text = f"""
### 📊 Dashboard KPIs
- **Total Candidates:** {kpis.get('total_candidates', 'N/A')}
- **Flagged for Audit:** {kpis.get('flagged_candidates', 'N/A')} ({kpis.get('flag_rate_pct', 'N/A')}%)
- **Mean Hiring Score:** {kpis.get('mean_hiring_score', 'N/A')}
- **Bias Flagged Candidates:** {kpis.get('bias_flagged', 'N/A')}
"""

with gr.Blocks(title="AI Hiring Decision Auditor") as demo:
    gr.Markdown("# 🤖 AI Hiring Decision Auditor Dashboard")
    gr.Markdown("Internal tool for HR compliance officers to audit automated hiring decisions.")
    gr.Markdown(kpi_text)

    with gr.Tabs():
        with gr.Tab("📈 Score Distribution by Gender"):
            gr.Image("score_distribution_by_gender.png")

        with gr.Tab("🔥 Bias Heatmap"):
            gr.Image("bias_heatmap_gender_education.png")
            gr.Markdown("Red cells indicate subgroups scoring significantly below average — primary bias risk zones.")

        with gr.Tab("⚠️ Risk Distribution by Gender"):
            gr.Image("risk_distribution_by_gender.png")

        with gr.Tab("💬 Sentiment vs Risk Label"):
            gr.Image("sentiment_vs_risk_label.png")
            gr.Markdown("Contradictions between recruiter sentiment and automated risk label indicate fail-silent risks.")

        with gr.Tab("📋 Subgroup Score Analysis"):
            gr.Dataframe(subgroup_df)

        with gr.Tab("🗂️ Full Audit Results"):
            gr.Dataframe(audit_df[['gender', 'education_level', 'hiring_score',
                                   'risk_label', 'note_sentiment', 'audit_result']].head(100))

demo.launch()