AKKI-AFK/ECL-Risk-Analyzer

1# app.py2import streamlit as st3import pandas as pd4import matplotlib.pyplot as plt5import google.generativeai as genai6import json, os, re, time7from datetime import datetime8from sqlalchemy import create_engine, Column, Integer, String, Float, DateTime9from sqlalchemy.orm import declarative_base, sessionmaker10 11# ========== CONFIG ==========12st.set_page_config(page_title="ECL Decision Assistant", layout="wide")13GEN_API_KEY = os.getenv("GEMINI_API_KEY")14if GEN_API_KEY:15    genai.configure(api_key=GEN_API_KEY)16else:17    st.warning("GEMINI_API_KEY not found in env. Set it in HF Space secrets to enable AI decisions.")18 19# Simple credential store (replace with secure store in production)20USERS = {21    "analyst": {"password": "analyst123", "role": "analyst"},22    "cro": {"password": "cro123", "role": "cro"},23}24 25# SQLite DB for persisting reports26DB_FILE = "reports.db"27engine = create_engine(f"sqlite:///{DB_FILE}", connect_args={"check_same_thread": False})28Base = declarative_base()29SessionLocal = sessionmaker(bind=engine)30 31class Report(Base):32    __tablename__ = "reports"33    id = Column(Integer, primary_key=True, index=True)34    segment = Column(String)35    pd = Column(Float)36    lgd = Column(Float)37    ead = Column(Float)38    ecl = Column(Float)39    action = Column(String)40    rationale = Column(String)41    confidence = Column(Float)42    generated_by = Column(String)43    created_at = Column(DateTime)44 45Base.metadata.create_all(bind=engine)46 47# ========== UTILITIES ==========48@st.cache_data49def process_loan_data(df: pd.DataFrame, segment_col: str = "loan_intent"):50    """Compute PD, LGD, EAD, ECL by segment column."""51    required = [segment_col, "credit_score", "loan_amnt", "loan_status"]52    df = df.dropna(subset=required)53    # ensure types54    df["loan_status"] = df["loan_status"].astype(int)55    df["credit_score"] = df["credit_score"].astype(float)56    df["loan_amnt"] = df["loan_amnt"].astype(float)57    group = df.groupby(segment_col)58    pd_seg = group["loan_status"].mean()59    lgd_seg = (1 - group["credit_score"].mean() / 850).clip(lower=0.0)60    ead_seg = group["loan_amnt"].sum()61    ecl_seg = pd_seg * lgd_seg * ead_seg62    ecl_df = pd.concat([pd_seg, lgd_seg, ead_seg, ecl_seg], axis=1)63    ecl_df.columns = ["PD", "LGD", "EAD", "ECL"]64    ecl_df = ecl_df.reset_index().rename(columns={segment_col: "segment"})65    return ecl_df66 67def sanitize_parse_json(text: str):68    """Extract first JSON object in text and parse it."""69    if not text:70        raise ValueError("Empty response")71    # remove common markdown fences72    text = re.sub(r"^```json\s*", "", text, flags=re.IGNORECASE)73    text = re.sub(r"^```\s*", "", text)74    text = re.sub(r"```$", "", text)75    # find JSON block76    m = re.search(r"\{.*\}", text, flags=re.DOTALL)77    if m:78        text = m.group(0)79    # attempt load80    return json.loads(text)81 82def get_gemini_decision_single(segment, pd_val, lgd_val, ead_val, ecl_val):83    """Single Gemini call per selected segment. Robust cleaning. Returns dict."""84    # If API key missing, return deterministic fallback85    if not GEN_API_KEY:86        return {"action": "maintain", "rationale": "No API key configured", "confidence": 0.0}87 88    prompt = f"""89You are a financial risk advisor. Return ONLY one valid JSON object with this schema:90{{"action":"increase_interest"|"reduce_disbursement"|"maintain","rationale":"string","confidence":float}}91 92Segment: {segment}93PD: {pd_val:.3f}94LGD: {lgd_val:.3f}95EAD: {ead_val:,.0f}96ECL: {ecl_val:,.0f}97 98Rules:99- PD > 0.25 => increase_interest100- 0.20 <= PD <= 0.25 => reduce_disbursement101- PD < 0.15 => maintain102 103Respond with a single JSON object and nothing else.104"""105 106    # Use model.generate_content with single prompt string (compat for HF)107    try:108        model = genai.GenerativeModel("gemini-2.5-flash-lite")109        resp = model.generate_content(prompt, generation_config={"temperature": 0.05})110        raw = resp.text if hasattr(resp, "text") else str(resp)111        # parse112        data = sanitize_parse_json(raw)113        # validate keys114        for k in ("action", "rationale", "confidence"):115            if k not in data:116                raise ValueError(f"Missing key: {k}")117        return data118    except Exception as e:119        # handle rate limits explicitly120        msg = str(e)121        if "429" in msg or "Resource exhausted" in msg:122            return {"action": "maintain", "rationale": "API quota exhausted - retry later", "confidence": 0.0}123        # fallback deterministic rule as final fallback124        if pd_val > 0.25:125            return {"action": "increase_interest", "rationale": "PD > 0.25 (deterministic fallback)", "confidence": 0.8}126        if 0.20 <= pd_val <= 0.25:127            return {"action": "reduce_disbursement", "rationale": "PD in 0.20-0.25 (deterministic fallback)", "confidence": 0.7}128        return {"action": "maintain", "rationale": "Fallback - parse or API error", "confidence": 0.0}129 130def save_report_to_db(row, decision, username):131    s = SessionLocal()132    r = Report(133        segment=row["segment"],134        pd=float(row["PD"]),135        lgd=float(row["LGD"]),136        ead=float(row["EAD"]),137        ecl=float(row["ECL"]),138        action=decision.get("action"),139        rationale=decision.get("rationale"),140        confidence=float(decision.get("confidence", 0.0)),141        generated_by=username,142        created_at=datetime.utcnow()143    )144    s.add(r)145    s.commit()146    s.refresh(r)147    s.close()148    return r.id149 150def load_reports_from_db(username, role):151    s = SessionLocal()152    if role == "cro":153        rows = s.query(Report).order_by(Report.created_at.desc()).all()154    else:155        rows = s.query(Report).filter(Report.generated_by == username).order_by(Report.created_at.desc()).all()156    df = pd.DataFrame([{157        "id": r.id,158        "segment": r.segment,159        "pd": r.pd,160        "lgd": r.lgd,161        "ead": r.ead,162        "ecl": r.ecl,163        "action": r.action,164        "rationale": r.rationale,165        "confidence": r.confidence,166        "generated_by": r.generated_by,167        "created_at": r.created_at168    } for r in rows])169    s.close()170    return df171 172# ========== UI - AUTH ==========173st.sidebar.title("Login")174username = st.sidebar.text_input("Username")175password = st.sidebar.text_input("Password", type="password")176if "auth_ok" not in st.session_state:177    st.session_state.auth_ok = False178if st.sidebar.button("Sign in"):179    user = USERS.get(username)180    if user and user["password"] == password:181        st.session_state.auth_ok = True182        st.session_state.username = username183        st.session_state.role = user["role"]184        st.sidebar.success(f"Signed in as {username} ({user['role']})")185    else:186        st.sidebar.error("Invalid credentials")187if not st.session_state.auth_ok:188    st.stop()189 190# ========== MAIN ==========191st.header("ECL Decision Assistant")192st.write(f"Signed in as **{st.session_state.username}** ({st.session_state.role})")193 194# Upload CSV195uploaded = st.file_uploader("Upload loan CSV (must contain loan_intent, credit_score, loan_amnt, loan_status)", type=["csv"])196if uploaded:197    df = pd.read_csv(uploaded)198    st.write("Sample rows:")199    st.dataframe(df.head(), width='stretch')200    # allow user to choose segmentation column201    seg_col = st.selectbox("Segment by column", options=[c for c in df.columns if df[c].dtype == object] , index=0)202    ecl_df = process_loan_data(df, segment_col=seg_col)203    st.subheader("Segment-level ECL Summary")204    st.dataframe(ecl_df, width='stretch')205 206    # Plots207    col1, col2 = st.columns(2)208    with col1:209        st.subheader("ECL by Segment")210        fig, ax = plt.subplots(figsize=(8, 3))211        ax.bar(ecl_df["segment"], ecl_df["ECL"])212        ax.set_xlabel("Segment"); ax.set_ylabel("ECL"); plt.xticks(rotation=45)213        st.pyplot(fig)214    with col2:215        st.subheader("PD by Segment")216        fig2, ax2 = plt.subplots(figsize=(8, 3))217        ax2.bar(ecl_df["segment"], ecl_df["PD"], color="gray")218        ax2.set_xlabel("Segment"); ax2.set_ylabel("PD"); plt.xticks(rotation=45)219        st.pyplot(fig2)220 221    # Select single segment for Gemini222    st.subheader("Analyze one segment (single API call)")223    selected = st.selectbox("Choose a segment to analyze", ecl_df["segment"].tolist())224    row = ecl_df[ecl_df["segment"] == selected].iloc[0]225    st.write(f"PD: {row.PD:.3f} | LGD: {row.LGD:.3f} | EAD: {row.EAD:,.0f} | ECL: {row.ECL:,.0f}")226 227    # Optionally show top segments only to reduce API usage228    max_n = len(ecl_df)229    default_n = min(5, max_n)230    top_n = st.number_input("Show top N segments by ECL (for reference)", min_value=1, max_value=max_n, value=default_n)231    st.write(ecl_df.sort_values("ECL", ascending=False).head(top_n))232 233    if st.button("Request Gemini decision for selected segment"):234        with st.spinner("Querying Gemini (single call)..."):235            decision = get_gemini_decision_single(row["segment"], row["PD"], row["LGD"], row["EAD"], row["ECL"])236            # save237            rec_id = save_report_to_db(row, decision, st.session_state.username)238        st.success("Decision recorded")239        st.json({"record_id": rec_id, "segment": row["segment"], "decision": decision})240 241# Historical reports section242st.subheader("Past Reports")243reports_df = load_reports_from_db(st.session_state.username, st.session_state.role)244if not reports_df.empty:245    st.dataframe(reports_df, width='stretch')246    # allow filtering by action247    action_filter = st.selectbox("Filter by action (All / increase_interest / reduce_disbursement / maintain)", ["All", "increase_interest", "reduce_disbursement", "maintain"])248    if action_filter != "All":249        st.dataframe(reports_df[reports_df["action"] == action_filter], width='stretch')250    if st.button("Download reports CSV"):251        st.download_button("Download", reports_df.to_csv(index=False).encode("utf-8"), file_name="reports.csv", mime="text/csv")252else:253    st.info("No reports recorded yet (use 'Request Gemini decision' to create one).")