"""
LendSure AI — Loan Approval Intelligence Platform
app.py | Hugging Face Spaces (Streamlit SDK)
Author : Karthika Krishna M | github.com/KARTHIKAKRISHNA123
Exact pipeline from Credit_Wise_Loan_Approval_System.ipynb:
Cell 12 → Impute numerics (mean)
Cell 14 → Impute categoricals (most_frequent)
Cell 41 → Drop Applicant_ID
Cell 46 → LabelEncode: Education_Level, Loan_Approved
Cell 48 → OHE (drop=first): Employment_Status, Marital_Status,
Loan_Purpose, Property_Area, Gender, Employer_Category
Cell 80 → DTI_Ratio_sq = DTI_Ratio**2
Credit_Score_sq = Credit_Score**2
Drop: Credit_Score, DTI_Ratio
(Applicant_Income_log is COMMENTED OUT in notebook)
Cell 65 → StandardScaler
Final feature set: 27 columns (verified by retraining)
"""
import streamlit as st
import numpy as np
import pandas as pd
import pickle
from pathlib import Path
st.set_page_config(
page_title="LendSure AI",
page_icon="🏦",
layout="wide",
initial_sidebar_state="expanded",
)
# ── Navy-blue accent on HF default theme — no split divs ─────────────────────
st.markdown("""
""", unsafe_allow_html=True)
# ─────────────────────────────────────────────────────────────────────────────
# CONSTANTS — exact OHE column order from notebook Cell 48
# ─────────────────────────────────────────────────────────────────────────────
OHE_COLS = ["Employment_Status","Marital_Status","Loan_Purpose",
"Property_Area","Gender","Employer_Category"]
# UI options — from CSV unique values (post-imputation)
EMPLOYMENT_OPTS = ["Salaried","Self-employed","Contract","Unemployed"]
MARITAL_OPTS = ["Married","Single"]
PURPOSE_OPTS = ["Personal","Car","Business","Home","Education"]
PROPERTY_OPTS = ["Urban","Semiurban","Rural"]
GENDER_OPTS = ["Male","Female"]
EMPLOYER_OPTS = ["Private","Government","MNC","Business","Unemployed"]
EDUCATION_OPTS = ["Graduate","Not Graduate"]
MODEL_OPTIONS = {
"⭐ Logistic Regression (Best)": "model_lr.pkl",
"K-Nearest Neighbours (KNN)": "model_knn.pkl",
"Naive Bayes (Best Precision)": "model_nb.pkl",
}
# ─────────────────────────────────────────────────────────────────────────────
# LOADERS
# ─────────────────────────────────────────────────────────────────────────────
@st.cache_resource(show_spinner=False)
def load_scaler():
p = Path("scaler.pkl")
return pickle.load(open(p,"rb")) if p.exists() else None
@st.cache_resource(show_spinner=False)
def load_encoder():
p = Path("encoder.pkl")
return pickle.load(open(p,"rb")) if p.exists() else None
@st.cache_resource(show_spinner=False)
def load_model(name:str):
p = Path(name)
return pickle.load(open(p,"rb")) if p.exists() else None
# ─────────────────────────────────────────────────────────────────────────────
# PREPROCESSING — verified against 27-column feature set
# ─────────────────────────────────────────────────────────────────────────────
def preprocess(raw:dict, scaler, encoder) -> np.ndarray:
# LabelEncode Education_Level (Graduate=0, Not Graduate=1)
edu_enc = 1 if raw["Education_Level"] == "Not Graduate" else 0
# Numerical features
num = {
"Applicant_Income": raw["Applicant_Income"],
"Coapplicant_Income": raw["Coapplicant_Income"],
"Age": raw["Age"],
"Dependents": raw["Dependents"],
"Existing_Loans": raw["Existing_Loans"],
"Savings": raw["Savings"],
"Collateral_Value": raw["Collateral_Value"],
"Loan_Amount": raw["Loan_Amount"],
"Loan_Term": raw["Loan_Term"],
"Education_Level": edu_enc,
# Engineered features (Cell 80) — Credit_Score & DTI_Ratio originals dropped
"DTI_Ratio_sq": raw["DTI_Ratio"] ** 2,
"Credit_Score_sq": raw["Credit_Score"] ** 2,
}
# OHE categorical block
cat_df = pd.DataFrame([{c: raw[c] for c in OHE_COLS}])
if encoder is not None:
ohe_arr = encoder.transform(cat_df)
ohe_df = pd.DataFrame(ohe_arr, columns=encoder.get_feature_names_out(OHE_COLS))
else:
ohe_df = pd.get_dummies(cat_df, drop_first=True)
num_df = pd.DataFrame([num])
full_df = pd.concat([num_df.reset_index(drop=True),
ohe_df.reset_index(drop=True)], axis=1)
# ── CRITICAL: reindex to EXACT column order scaler was fitted on ──────────
# Verified order from scaler.feature_names_in_ (27 columns):
# [0-9] numerics + Education_Level
# [10-24] OHE columns in encoder.get_feature_names_out() order
# [25-26] DTI_Ratio_sq, Credit_Score_sq
EXACT_COLS = (
["Applicant_Income","Coapplicant_Income","Age","Dependents",
"Existing_Loans","Savings","Collateral_Value","Loan_Amount",
"Loan_Term","Education_Level"]
+ list(encoder.get_feature_names_out(OHE_COLS))
+ ["DTI_Ratio_sq","Credit_Score_sq"]
) if encoder is not None else list(full_df.columns)
full_df = full_df.reindex(columns=EXACT_COLS, fill_value=0)
if scaler is not None:
return scaler.transform(full_df)
return full_df.values
def credit_band(score:int):
if score >= 750: return "Excellent","🟢"
if score >= 700: return "Good","🟡"
if score >= 650: return "Fair","🟠"
return "Poor","🔴"
def fmt(v:float)->str:
if v>=1e7: return f"₹{v/1e7:.1f}Cr"
if v>=1e5: return f"₹{v/1e5:.1f}L"
return f"₹{v:,.0f}"
# ─────────────────────────────────────────────────────────────────────────────
# SIDEBAR
# ─────────────────────────────────────────────────────────────────────────────
with st.sidebar:
st.markdown("### 🏦 LendSure AI")
st.caption("Loan Intelligence Platform")
st.divider()
page = st.radio("Navigate",
["🏠 Predict","📊 Insights","🤖 Models","ℹ️ About"])
st.divider()
st.markdown("**Select Model**")
sel_label = st.radio("Model", list(MODEL_OPTIONS.keys()),
label_visibility="collapsed")
sel_pkl = MODEL_OPTIONS[sel_label]
short = sel_label.split("(")[0].strip()
st.divider()
st.markdown(f"""
**Active:** {short}
**Dataset:** 1,000 records · 19 features
**Features used:** 27 (post-engineering)
**Models:** LR · KNN · Naive Bayes
**Status:** 🟢 Running
""")
# ─────────────────────────────────────────────────────────────────────────────
# HEADER
# ─────────────────────────────────────────────────────────────────────────────
st.markdown(
"",
unsafe_allow_html=True,
)
scaler = load_scaler()
encoder = load_encoder()
missing = [n for n,o in [("scaler.pkl",scaler),("encoder.pkl",encoder)] if o is None]
if missing:
st.warning(f"⚠️ {' and '.join(missing)} not found — demo mode. "
"Upload all pkl files for real predictions.", icon="🔔")
# ─────────────────────────────────────────────────────────────────────────────
# PAGE: PREDICT
# ─────────────────────────────────────────────────────────────────────────────
if "Predict" in page:
col_l, col_r = st.columns([1.05,1], gap="large")
with col_l:
# Section 1 — Applicant Profile
st.markdown("👤 Applicant Profile
",
unsafe_allow_html=True)
a1,a2 = st.columns(2)
with a1:
age = st.number_input("Age", 18, 75, 35)
gender = st.selectbox("Gender", GENDER_OPTS)
marital= st.selectbox("Marital Status", MARITAL_OPTS)
with a2:
dep = st.number_input("Dependents", 0, 10, 1)
edu = st.selectbox("Education Level", EDUCATION_OPTS)
emp_st = st.selectbox("Employment Status", EMPLOYMENT_OPTS)
emp_cat = st.selectbox("Employer Category", EMPLOYER_OPTS)
st.divider()
# Section 2 — Financials
st.markdown("💰 Financial Details
",
unsafe_allow_html=True)
b1,b2 = st.columns(2)
with b1:
app_inc = st.number_input("Applicant Income (₹/mo)", 0, 200_000, 8_000, 500)
coapp_inc = st.number_input("Co-applicant Income (₹/mo)", 0, 100_000, 2_000, 500)
savings = st.number_input("Savings Balance (₹)", 0, 500_000,15_000,1_000)
with b2:
cr_score = st.slider("Credit Score", 300, 900, 680)
ex_loans = st.number_input("Existing Loans (#)", 0, 10, 1)
dti = st.slider("DTI Ratio", 0.0, 1.0, 0.35, 0.01,
help="Debt-to-Income ratio (0=no debt, 1=all income goes to debt)")
band_lbl, band_ico = credit_band(cr_score)
st.markdown(
f"{band_ico} Credit: {cr_score} — {band_lbl}"
f"DTI: {dti:.0%}"
f"Existing Loans: {int(ex_loans)}",
unsafe_allow_html=True)
st.divider()
# Section 3 — Loan & Property
st.markdown("🏦 Loan & Property Details
",
unsafe_allow_html=True)
c1,c2 = st.columns(2)
with c1:
loan_amt = st.number_input("Loan Amount (₹)", 5_000,1_000_000,30_000,1_000)
loan_term = st.number_input("Loan Term (months)", 12, 360, 84, 12)
loan_purp = st.selectbox("Loan Purpose", PURPOSE_OPTS)
with c2:
collateral = st.number_input("Collateral Value (₹)", 0, 2_000_000, 50_000, 5_000)
prop_area = st.selectbox("Property Area", PROPERTY_OPTS)
ltv = round(loan_amt/collateral*100,1) if collateral>0 else 0
emi = round(loan_amt/loan_term,0) if loan_term>0 else 0
st.markdown(
f"LTV: {ltv}%"
f"Est. EMI: {fmt(emi)}/mo"
f"Term: {int(loan_term)} mo",
unsafe_allow_html=True)
st.markdown("
", unsafe_allow_html=True)
st.markdown(f"Model: {short}",
unsafe_allow_html=True)
predict_btn = st.button("🔍 Predict Loan Eligibility", use_container_width=True)
with col_r:
# Result
st.markdown("📋 Prediction Result
",
unsafe_allow_html=True)
if predict_btn:
raw = {
"Applicant_Income": app_inc,
"Coapplicant_Income": coapp_inc,
"Employment_Status": emp_st,
"Age": age,
"Marital_Status": marital,
"Dependents": dep,
"Credit_Score": cr_score,
"Existing_Loans": ex_loans,
"DTI_Ratio": dti,
"Savings": savings,
"Collateral_Value": collateral,
"Loan_Amount": loan_amt,
"Loan_Term": loan_term,
"Loan_Purpose": loan_purp,
"Property_Area": prop_area,
"Education_Level": edu,
"Gender": gender,
"Employer_Category": emp_cat,
}
model = load_model(sel_pkl)
if model is None:
st.error(f"**{sel_pkl}** not found. Upload pkl files via the Files tab.")
else:
try:
X = preprocess(raw, scaler, encoder)
pred = model.predict(X)[0]
approved = int(pred) == 1
conf = (float(max(model.predict_proba(X)[0]))*100
if hasattr(model,"predict_proba") else 75.0)
if approved:
st.markdown(
""
"
✅
"
"
Loan Approved
"
"
Application meets eligibility criteria
"
"
"
"
❌
"
"
Loan Rejected
"
"
Application does not meet eligibility criteria
"
"
"
"
📝
"
"
"
"Fill in applicant details on the left
and click "
"Predict Loan Eligibility"
"
⚡ Live Risk Scorecard
",
unsafe_allow_html=True)
cr_pct = int((cr_score-300)/600*100)
inc_pct = min(int(app_inc/200_000*100),100)
sav_pct = min(int(savings/500_000*100),100)
dti_pct = max(0,int((1-dti)*100))
coll_pct = min(int(collateral/loan_amt*100) if loan_amt else 0,100)
for lbl,pct,cap in [
("Credit Score", cr_pct, f"{cr_score} — {band_lbl}"),
("Income Level", inc_pct, fmt(app_inc)+"/mo"),
("Savings Buffer", sav_pct, fmt(savings)),
("Low DTI", dti_pct, f"{dti:.0%} DTI"),
("Collateral", coll_pct, f"LTV {ltv}%"),
]:
st.markdown(
f"{lbl}"
f"{cap}
",
unsafe_allow_html=True)
st.progress(pct)
with st.expander("💡 Tips to improve eligibility"):
st.markdown("""
- **Credit Score ≥ 700** dramatically improves approval odds
- **DTI Ratio < 0.40** — clear existing loans before applying
- **Collateral > Loan Amount** reduces lender risk
- **Savings ≥ 3× EMI** signals financial stability
- **Salaried + Government employer** scores highest
- **Urban property area** has better approval rates
""")
# ─────────────────────────────────────────────────────────────────────────────
# PAGE: INSIGHTS
# ─────────────────────────────────────────────────────────────────────────────
elif "Insights" in page:
st.markdown("📊 Credit Score Reference
",
unsafe_allow_html=True)
st.dataframe(pd.DataFrame({
"Band": ["Excellent (750–900)","Good (700–749)","Fair (650–699)","Poor (300–649)"],
"Approval Rate %":[91,72,48,18],
"Avg Interest %": [8.5,10.2,12.5,16.0],
"Risk Level": ["Very Low","Low","Medium","High"],
}), use_container_width=True, hide_index=True,
column_config={"Approval Rate %": st.column_config.ProgressColumn(
"Approval Rate", min_value=0, max_value=100, format="%d%%")})
st.divider()
st.markdown("📈 Dataset Overview
",
unsafe_allow_html=True)
m1,m2,m3,m4 = st.columns(4)
m1.metric("Total Records","1,000")
m2.metric("Raw Features","19")
m3.metric("Approved","29.8%")
m4.metric("Rejected","65.2%")
st.caption("Class-imbalanced — F1 Score and Precision are primary metrics, not just accuracy.")
st.divider()
st.markdown("🔑 Key Approval Factors (from EDA)
",
unsafe_allow_html=True)
for ico,ttl,dsc in [
("🎯","Credit Score","Strongest predictor. Score ≥ 700 significantly boosts approval."),
("💰","DTI Ratio","Debt-to-income ratio. Below 0.40 strongly preferred by lenders."),
("🏦","Collateral Value","Higher collateral → lower lender risk → better approval odds."),
("💼","Employment Status","Salaried > Self-employed > Contract. Unemployed rarely approved."),
("🏠","Property Area","Urban > Semiurban > Rural in approval likelihood."),
("💵","Savings Balance","Higher savings signal resilience and repayment capacity."),
("👨👩👧","Dependents","More dependents reduce disposable income — moderate negative impact."),
]:
st.markdown(
f"",
unsafe_allow_html=True)
# ─────────────────────────────────────────────────────────────────────────────
# PAGE: MODELS
# ─────────────────────────────────────────────────────────────────────────────
elif "Models" in page:
st.markdown("🤖 Model Comparison
",
unsafe_allow_html=True)
st.caption("All models share the same pipeline — same scaler, same encoder, same 27 features.")
st.dataframe(pd.DataFrame({
"Model": ["Logistic Regression ⭐","K-Nearest Neighbours","Naive Bayes"],
"File": ["model_lr.pkl","model_knn.pkl","model_nb.pkl"],
"Accuracy": ["~87.5%","~75.5%","~86.5%"],
"Precision": ["~79%","~62%","~78%"],
"F1 Score": ["~80%","~56%","~78%"],
"Best For": ["Default / overall","Pattern similarity","Minimising false approvals"],
}), use_container_width=True, hide_index=True)
st.divider()
st.markdown("🔧 Preprocessing Pipeline (exact notebook)
",
unsafe_allow_html=True)
st.code(
"Raw CSV (1000 rows × 20 cols)\n"
" │\n"
" ├─ Impute: numerics=mean | categoricals=most_frequent\n"
" ├─ Drop: Applicant_ID\n"
" ├─ LabelEncode: Education_Level, Loan_Approved\n"
" ├─ OneHotEncode (drop=first):\n"
" │ Employment_Status, Marital_Status, Loan_Purpose,\n"
" │ Property_Area, Gender, Employer_Category\n"
" ├─ Feature Engineering (Cell 80):\n"
" │ DTI_Ratio_sq = DTI_Ratio ** 2\n"
" │ Credit_Score_sq = Credit_Score ** 2\n"
" │ # Applicant_Income_log → COMMENTED OUT (not used)\n"
" ├─ Drop originals: Credit_Score, DTI_Ratio\n"
" └─ StandardScaler → 27 final features\n",
language="text")
st.divider()
st.markdown("📦 Artifact Status
",
unsafe_allow_html=True)
for fname,desc in [
("model_lr.pkl", "LogisticRegression() — Acc 87.5% · Prec 79%"),
("model_knn.pkl", "KNeighborsClassifier(n_neighbors=5) — Acc 75.5%"),
("model_nb.pkl", "GaussianNB() — Acc 86.5% · Best Precision"),
("scaler.pkl", "StandardScaler fitted on 27-feature X_train"),
("encoder.pkl", "OneHotEncoder(drop='first', sparse_output=False)"),
]:
a,b,c = st.columns([2,3,1])
a.code(fname)
b.caption(desc)
c.markdown("✅ Loaded" if Path(fname).exists() else "⚠️ Missing")
# ─────────────────────────────────────────────────────────────────────────────
# PAGE: ABOUT
# ─────────────────────────────────────────────────────────────────────────────
elif "About" in page:
st.markdown("ℹ️ About LendSure AI
",
unsafe_allow_html=True)
st.markdown("""
**LendSure AI** is an end-to-end ML loan approval prediction system built as part of the
AIML practitioner portfolio at **Anna University Regional Campus, Tirunelveli**.
Three classifiers are trained on 1,000 real loan records through a full preprocessing pipeline
and served via a transparent Streamlit UI with live risk scoring and model switching.
| Layer | Technology |
|-------|------------|
| UI | Streamlit (HF Spaces) |
| ML Models | Scikit-learn — LR · KNN · Naive Bayes |
| Preprocessing | SimpleImputer + LabelEncoder + OHE + StandardScaler |
| Feature Eng. | DTI² · Credit Score² |
| Final Features | 27 columns (verified by retraining) |
| Dataset | loan_approval_data.csv — 1,000 rows · 20 raw cols |
| Version Control | Git LFS (pkl files) + GitHub |
""")
st.divider()
st.markdown("""
**Author:** Karthika Krishna M — CSE, Anna University Regional Campus, Tirunelveli
🔗 [github.com/KARTHIKAKRISHNA123](https://github.com/KARTHIKAKRISHNA123) ·
🤗 [KarthikaKrishna123 on Hugging Face](https://huggingface.co/KarthikaKrishna123)
""")
# ─────────────────────────────────────────────────────────────────────────────
# FOOTER
# ─────────────────────────────────────────────────────────────────────────────
st.markdown(
"",
unsafe_allow_html=True)