दिन 60: Project 5: App Deploy और Final Report

import numpy as np
import pandas as pd

def load_churn(path="telecom_churn.csv"):
    # "None" means no internet plan – stop pandas from turning it into NaN
    df = pd.read_csv(path, keep_default_na=False, na_values=[""])
    df["total_charges"] = df["total_charges"].fillna(df["monthly_charges"] * df["tenure_months"])
    df["churned"] = (df["churn"] == "Yes").astype(int)
    return df

def add_features(df):
    df = df.copy()
    df["tenure_band"] = pd.cut(df["tenure_months"], [0, 6, 12, 24, 48, 72], labels=["0-6", "7-12", "13-24", "25-48", "49-72"]).astype(str)
    df["is_new"] = (df["tenure_months"] <= 6).astype(int)
    df["has_complaint"] = (df["complaints"] > 0).astype(int)
    df["charge_per_gb"] = (df["monthly_charges"] / df["data_usage_gb"].clip(lower=1)).round(2)
    df["avg_monthly_paid"] = (df["total_charges"] / df["tenure_months"]).round(1)
    df["monthly_no_support"] = ((df["contract"] == "Monthly") & (df["tech_support"] == "No")).astype(int)
    return df

# original columns – the champion model uses these
NUM = ["senior_citizen", "tenure_months", "monthly_charges", "total_charges", "complaints", "data_usage_gb"]
CAT = ["gender", "contract", "internet_service", "tech_support", "payment_method"]
# original + engineered columns (tested on Day 58)
NUM_ALL = NUM + ["is_new", "has_complaint", "charge_per_gb", "avg_monthly_paid", "monthly_no_support"]
CAT_ALL = CAT + ["tenure_band"]

import json
import joblib
import sklearn
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OneHotEncoder, StandardScaler
from sklearn.pipeline import Pipeline
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import cross_val_score, StratifiedKFold

df = load_churn()
X, y = df[NUM + CAT], df["churned"]
model = Pipeline([("prep", ColumnTransformer([("num", StandardScaler(), NUM), ("cat", OneHotEncoder(handle_unknown="ignore"), CAT)])),
                  ("clf", LogisticRegression(C=0.1, max_iter=5000))])
cv_auc = cross_val_score(model, X, y, cv=StratifiedKFold(5, shuffle=True, random_state=0), scoring="roc_auc")
model.fit(X, y)                                    # final model uses ALL 3,000 customers
joblib.dump(model, "churn_model_v1.joblib")
card = {"name": "TeleOne churn model", "version": "1.0.0", "algorithm": "logistic regression (C=0.1)",
        "trained_on": f"{len(X)} customers", "features": NUM + CAT,
        "cv_roc_auc": f"{cv_auc.mean():.3f} ± {cv_auc.std():.3f}", "locked_test_roc_auc_day59": 0.832,
        "intended_use": "rank customers for retention calls; not for pricing or credit decisions",
        "monitoring": "monthly PSI on inputs, quarterly AUC on new outcomes, retrain if AUC < 0.78",
        "sklearn_version": sklearn.__version__}
json.dump(card, open("model_card.json", "w", encoding="utf-8"), indent=2, ensure_ascii=False)
print(json.dumps(card, indent=2, ensure_ascii=False))

import joblib
import numpy as np
import pandas as pd

def load_churn(path="telecom_churn.csv"):
    df = pd.read_csv(path, keep_default_na=False, na_values=[""])
    df["total_charges"] = df["total_charges"].fillna(df["monthly_charges"] * df["tenure_months"])
    return df

model = joblib.load("churn_model_v1.joblib")
pre, clf = model[0], model[-1]
names = pre.get_feature_names_out()

ACTIONS = {"contract_Monthly": "Offer yearly plan at 15% off", "tech_support_No": "Add 3 months free tech support",
           "complaints": "Senior agent callback on open complaints", "tenure_months": "Welcome-back / onboarding help",
           "monthly_charges": "Review plan – suggest a better-value pack", "internet_service_Fiber": "Fiber quality check visit"}

def label(j, Z, mean):
    key = names[j].split("__", 1)[1]
    if names[j].startswith("num__"):
        if key == "senior_citizen":
            return key, "senior citizen"
        return key, ("high " if Z[j] > mean[j] else "short " if key == "tenure_months" else "low ") + key.replace("_", " ")
    return key, key.replace("_", " ", 1).replace("_", " ")

def explain(Z, mean):
    contrib = (Z - mean) * clf.coef_[0]
    top = [label(j, Z, mean) for j in np.argsort(contrib)[::-1][:3] if contrib[j] > 0]
    action = next((ACTIONS[k] for k, _ in top if k in ACTIONS), "Standard loyalty offer")
    return ", ".join(t for _, t in top), action

def monthly_risk_list(customers, capacity=600):
    X = customers[model.feature_names_in_]
    p = model.predict_proba(X)[:, 1]
    Z = pre.transform(X); Z = Z.toarray() if hasattr(Z, "toarray") else Z
    mean = np.asarray(Z.mean(axis=0)).ravel()
    out = customers[["customer_id", "contract", "tenure_months", "monthly_charges"]].copy()
    out["churn_risk"] = p.round(3)
    out["risk_band"] = pd.cut(p, [0, 0.3, 0.6, 1], labels=["Low", "Medium", "High"])
    expl = [explain(Z[i], mean) for i in range(len(out))]
    out["top_reasons"] = [e[0] for e in expl]; out["suggested_action"] = [e[1] for e in expl]
    out["value_at_risk_₹"] = (out["churn_risk"] * out["monthly_charges"] * 12).round(0)
    return out.sort_values("value_at_risk_₹", ascending=False).head(capacity)

active = load_churn().query("churn == 'No'")          # this month: score customers who are still with us
risk = monthly_risk_list(active, capacity=600)
risk.to_csv("risk_list_this_month.csv", index=False)
print(f"Scored {len(active):,} active customers → call list of {len(risk)} saved to risk_list_this_month.csv")
print(risk.head(6).to_string(index=False))
print("\nRisk bands in the call list:", risk["risk_band"].value_counts().to_dict())
hm = risk[risk["risk_band"].isin(["High", "Medium"])]
print(f"High + Medium: {len(hm)} customers | avg risk {hm['churn_risk'].mean():.0%} | value at risk ₹{hm['value_at_risk_₹'].sum() / 1e5:.1f} lakh/yr")

# crontab -e
0 6 1 * *  cd /opt/teleone-churn && /usr/bin/python3 score_customers.py >> logs/score.log 2>&1

# app_retention.py – run: streamlit run app_retention.py
import joblib
import pandas as pd
import streamlit as st

st.set_page_config(page_title="TeleOne Retention", page_icon="📞", layout="wide")
model = joblib.load("churn_model_v1.joblib")
st.title("📞 TeleOne – Retention Assistant")
tab1, tab2 = st.tabs(["This month’s call list", "Check one customer"])

with tab1:
    risk = pd.read_csv("risk_list_this_month.csv")
    band = st.multiselect("Risk band", ["High", "Medium", "Low"], default=["High", "Medium"])
    view = risk[risk["risk_band"].isin(band)]
    c1, c2, c3 = st.columns(3)
    c1.metric("Customers to call", len(view))
    c2.metric("Avg risk", f"{view['churn_risk'].mean():.0%}")
    c3.metric("Value at risk", f"₹{view['value_at_risk_₹'].sum() / 1e5:.1f} lakh / yr")
    st.dataframe(view, use_container_width=True, hide_index=True)
    st.download_button("Download CSV for the dialler", view.to_csv(index=False), "call_list.csv")

with tab2:
    a, b = st.columns(2)
    row = {"senior_citizen": int(a.checkbox("Senior citizen")),
           "tenure_months": a.slider("Tenure (months)", 1, 72, 6),
           "monthly_charges": a.number_input("Monthly bill (₹)", 149, 1500, 850),
           "complaints": a.slider("Complaints (last 6 months)", 0, 8, 2),
           "data_usage_gb": a.number_input("Data usage (GB/month)", 0.0, 300.0, 40.0),
           "gender": b.selectbox("Gender", ["Female", "Male"]),
           "contract": b.selectbox("Contract", ["Monthly", "One year", "Two year"]),
           "internet_service": b.selectbox("Internet", ["Fiber", "DSL", "None"]),
           "tech_support": b.selectbox("Tech support", ["No", "Yes", "No internet"]),
           "payment_method": b.selectbox("Payment", ["UPI AutoPay", "Credit Card", "Bank Transfer", "Cash"])}
    row["total_charges"] = row["monthly_charges"] * row["tenure_months"]
    p = model.predict_proba(pd.DataFrame([row]))[0, 1]
    st.metric("Churn risk", f"{p:.0%}")
    st.progress(float(p))
    if p >= 0.6:
        st.error("High risk – call today with a retention offer")
    elif p >= 0.3:
        st.warning("Medium risk – include in this month’s campaign")
    else:
        st.success("Low risk – no action needed")

<div style="font-family:system-ui,sans-serif;padding:12px;background:#fff">
<h3 style="margin:0 0 8px">📞 TeleOne – Retention Assistant</h3>
<div style="display:flex;gap:6px;margin-bottom:10px"><span style="padding:5px 10px;border-bottom:2px solid #ff4b4b;font-weight:600">This month’s call list</span><span style="padding:5px 10px;color:#6b7280">Check one customer</span></div>
<div style="display:grid;grid-template-columns:repeat(3,1fr);gap:8px;margin-bottom:10px">
<div style="background:#f8fafc;border-radius:8px;padding:8px"><small>Customers to call</small><div style="font-size:22px;font-weight:700">463</div></div>
<div style="background:#f8fafc;border-radius:8px;padding:8px"><small>Avg risk</small><div style="font-size:22px;font-weight:700">49%</div></div>
<div style="background:#f8fafc;border-radius:8px;padding:8px"><small>Value at risk</small><div style="font-size:22px;font-weight:700">₹22.5 lakh / yr</div></div></div>
<table style="width:100%;border-collapse:collapse;font-size:12px"><tr style="background:#f1f5f9"><th align="left">customer</th><th>risk</th><th align="left">top reasons</th><th align="left">suggested action</th></tr>
<tr><td>CUST12993</td><td align="center"><b style="color:#dc2626">85%</b></td><td>high monthly charges, contract Monthly, short tenure</td><td>Review plan – better-value pack</td></tr>
<tr><td>CUST10745</td><td align="center"><b style="color:#dc2626">81%</b></td><td>high complaints, high monthly charges, contract Monthly</td><td>Senior agent callback on complaints</td></tr>
<tr><td>CUST10268</td><td align="center"><b style="color:#dc2626">80%</b></td><td>contract Monthly, high monthly charges, high complaints</td><td>Offer yearly plan at 15% off</td></tr></table>
<button style="margin-top:10px;background:#ff4b4b;color:#fff;border:0;padding:7px 12px;border-radius:8px">Download CSV for the dialler</button></div>

import joblib, pandas as pd
from flask import Flask, request, jsonify
app = Flask(__name__)
model = joblib.load("churn_model_v1.joblib")
FIELDS = list(model.feature_names_in_)

@app.post("/score")
def score():
    rows = request.get_json(silent=True)
    rows = rows if isinstance(rows, list) else [rows]
    bad = [i for i, r in enumerate(rows) if not isinstance(r, dict) or set(FIELDS) - set(r)]
    if bad:
        return jsonify(error=f"rows {bad} are missing fields", required=FIELDS), 400
    p = model.predict_proba(pd.DataFrame(rows)[FIELDS])[:, 1]
    return jsonify([{"churn_risk": round(float(x), 3), "call": bool(x >= 0.3)} for x in p])

client = app.test_client()
cust = {"senior_citizen": 0, "tenure_months": 3, "monthly_charges": 950, "total_charges": 2850, "complaints": 2,
        "data_usage_gb": 35.0, "gender": "Female", "contract": "Monthly", "internet_service": "Fiber",
        "tech_support": "No", "payment_method": "Cash"}
loyal = dict(cust, tenure_months=60, contract="Two year", tech_support="Yes", complaints=0, total_charges=57000)
r = client.post("/score", json=[cust, loyal]); print(r.status_code, r.get_json())
r = client.post("/score", json={"tenure_months": 5}); print(r.status_code, r.get_json()["error"])

teleone-churn/
├── README.md                ← business summary first (template below)
├── PROJECT_CHARTER.md       ← Day 56
├── requirements.txt
├── data/sample_customers.csv
├── notebooks/
│   ├── 01_problem_and_audit.ipynb      (Day 56)
│   ├── 02_eda_insights.ipynb           (Day 57)
│   ├── 03_features_models.ipynb        (Day 58)
│   └── 04_evaluation_impact.ipynb      (Day 59)
├── src/  churn.py  (load_churn, add_features, make_model)
├── train_final.py   score_customers.py   api.py   app_retention.py
├── models/  churn_model_v1.joblib   model_card.json
└── reports/ figures/   final_report.md   slides.pdf

# TeleOne Churn Prediction – Final Report

## 1. Problem
29% of customers churned, taking ≈ ₹6.6 lakh of monthly revenue (≈ ₹80 lakh a year).
The retention team can call ~600 customers a month and needs to know whom to call and why.

## 2. What drives churn (EDA, Day 57)
- Monthly contracts churn ~44% vs ~9% for two-year contracts (largest effect).
- First 6 months, multiple complaints, no tech support and Fiber plans raise risk.
- Gender has no effect; payment method differences were not significant.

## 3. Model (Days 58–59)
- Logistic regression on the 11 original columns (engineered features added nothing).
- 5-fold CV ROC-AUC ≈ 0.82; locked test ROC-AUC 0.83 – no sign of leakage or overfitting.
- Well calibrated; similar performance for women and men.

## 4. Business impact (agreed assumptions)
- Calling the top 20% reaches 2.3× more churners than random calling.
- Expected net ≈ ₹12 lakh per monthly campaign vs ≈ ₹3 lakh for random calls.

## 5. Recommendations
1. Use the monthly risk list (sorted by value at risk) for retention calls.
2. Bundle free tech support for monthly Fiber customers; push yearly-plan upgrades.
3. Run a 2-month A/B pilot (call vs hold-out) to measure the real save-rate.
4. Monitor input drift monthly; retrain when AUC on new outcomes < 0.78.

## 6. Limitations
Simulated data, single snapshot, save-rate and offer acceptance are assumptions.

“TeleOne was losing 29% of its customers – about ₹80 lakh a year – and the retention
team could only call 600 people a month. I framed it as a classification problem, audited
the data (and caught a silent NaN bug), and found in EDA that monthly contracts, the first
six months, complaints and missing tech support drive churn.
I compared five models with cross-validation; plain logistic regression won with AUC 0.83
on a locked test set, and my engineered features did not help, so I kept it simple.
Using agreed costs, calling the top 20% reaches 2.3× more churners than random calls –
about ₹12 lakh net per campaign. I shipped a monthly risk list with reasons and suggested
offers, a Streamlit app for agents and a Flask API, and proposed an A/B pilot to measure
the real save-rate.”