"""FraudSentinel Tier-1 inference helper.
Loads the trained LightGBM scorers and returns a risk score + a routing decision
(whether to escalate the case to the Tier-2 LLM for explanation).

Usage:
    from infer import CardScorer, AMLScorer
    cs = CardScorer()                 # loads cc_lgbm_model.txt + preproc
    out = cs.score(transaction_dict)  # -> {"risk": 0.97, "route_to_llm": True}
"""
import json
import numpy as np
import pandas as pd
import lightgbm as lgb
import joblib

# ---------------- Card fraud scorer ----------------
class CardScorer:
    def __init__(self, model="cc_lgbm_model.txt", preproc="cc_lgbm_preproc.joblib",
                 metrics="cc_lgbm_metrics.json"):
        self.model = lgb.Booster(model_file=model)
        pp = joblib.load(preproc)
        self.cat_p95 = pp["cat_p95"]; self.cat_rate = pp["cat_rate"]
        self.features = pp["features"]; self.cat_cols = pp["cat_cols"]
        self.threshold = json.load(open(metrics))["routing_threshold"]

    def _row(self, t):
        import math
        R = 6371.0
        def hav(a, b, c, d):
            p1, p2 = math.radians(a), math.radians(c)
            dphi = math.radians(c-a); dl = math.radians(d-b)
            x = math.sin(dphi/2)**2 + math.cos(p1)*math.cos(p2)*math.sin(dl/2)**2
            return 2*R*math.asin(min(1, math.sqrt(x)))
        dt = pd.to_datetime(t["trans_date_trans_time"])
        amt = float(t["amt"]); cat = t["category"]
        p95 = self.cat_p95.get(cat, np.median(list(self.cat_p95.values())))
        f = {
            "log_amt": np.log1p(amt), "amt": amt, "hour": dt.hour, "dow": dt.dayofweek,
            "is_night": int(dt.hour >= 22 or dt.hour <= 4),
            "age": (dt - pd.to_datetime(t["dob"])).days/365.25,
            "geo_km": hav(t["lat"], t["long"], t["merch_lat"], t["merch_long"]),
            "log_city_pop": np.log1p(t.get("city_pop", 0)),
            "mins_since_last": t.get("mins_since_last", 99999),
            "tx_24h": t.get("tx_24h", 0), "amt_24h": t.get("amt_24h", 0.0),
            "tx_1h": t.get("tx_1h", 0),
            "amt_over_p95": int(amt > p95), "amt_to_p95": amt/(p95+1e-6),
            "cat_fraud_rate": self.cat_rate.get(cat, 0.0),
            "category": cat, "gender": t.get("gender"), "state": t.get("state"),
        }
        df = pd.DataFrame([f])
        for c in self.cat_cols:
            df[c] = df[c].astype("category")
        return df[self.features]

    def score(self, t):
        r = float(self.model.predict(self._row(t))[0])
        return {"risk": r, "route_to_llm": r >= self.threshold, "tier": "card"}

# ---------------- AML scorer (recall-oriented pre-filter) ----------------
class AMLScorer:
    def __init__(self, model="aml_lgbm_model.txt", preproc="aml_lgbm_preproc.joblib",
                 metrics="aml_lgbm_metrics.json"):
        self.model = lgb.Booster(model_file=model)
        pp = joblib.load(preproc)
        self.g = pp["graph"]; self.features = pp["features"]; self.cat_cols = pp["cat_cols"]
        self.threshold = json.load(open(metrics))["routing_threshold"]

    def _row(self, t):
        ts = pd.to_datetime(t["Timestamp"], format="%Y/%m/%d %H:%M", errors="coerce")
        snd, rcv = t["Account"], t["Account.1"]
        paid = float(t["Amount Paid"]); recv = float(t["Amount Received"])
        gm = self.g
        f = {
            "hour": ts.hour if pd.notna(ts) else 0, "dow": ts.dayofweek if pd.notna(ts) else 0,
            "log_paid": np.log1p(paid), "log_recv": np.log1p(recv),
            "amt_diff": abs(paid-recv),
            "ccy_mismatch": int(t["Receiving Currency"] != t["Payment Currency"]),
            "self_loop": int(snd == rcv), "is_round": int(paid % 100 == 0),
            "same_bank": int(t["From Bank"] == t["To Bank"]),
            "snd_out_deg": gm["out_deg"].get(snd, 0), "snd_in_deg": gm["in_deg"].get(snd, 0),
            "snd_out_cnt": gm["out_cnt"].get(snd, 0), "snd_in_cnt": gm["in_cnt"].get(snd, 0),
            "snd_out_amt_mean": gm["out_amt_mean"].get(snd, 0.0),
            "rcv_out_deg": gm["out_deg"].get(rcv, 0), "rcv_in_deg": gm["in_deg"].get(rcv, 0),
            "rcv_in_cnt": gm["in_cnt"].get(rcv, 0),
            "gather_scatter": int(gm["in_deg"].get(snd, 0) >= 5 and gm["out_deg"].get(snd, 0) >= 5),
            "amt_to_snd_mean": paid/(gm["out_amt_mean"].get(snd, 0.0)+1e-6),
            "Receiving Currency": t["Receiving Currency"],
            "Payment Currency": t["Payment Currency"], "Payment Format": t["Payment Format"],
        }
        df = pd.DataFrame([f])
        for c in self.cat_cols:
            df[c] = df[c].astype("category")
        return df[self.features]

    def score(self, t):
        r = float(self.model.predict(self._row(t))[0])
        return {"risk": r, "route_to_llm": r >= self.threshold, "tier": "aml",
                "note": "recall-oriented pre-filter; use GNN for precision"}

if __name__ == "__main__":
    cs = CardScorer()
    demo = {"trans_date_trans_time": "2020-06-10 02:14:00", "amt": 980.0, "category": "shopping_net",
            "lat": 40.0, "long": -75.0, "merch_lat": 48.0, "merch_long": 2.0, "dob": "1985-01-01",
            "city_pop": 50000, "gender": "M", "state": "PA", "tx_24h": 5, "amt_24h": 3200.0,
            "tx_1h": 3, "mins_since_last": 4}
    print("CARD demo:", cs.score(demo))