diff --git a/docker-compose.yml b/docker-compose.yml
index 735a22d..7cade23 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -130,6 +130,7 @@ services:
   #   depends_on:
   #     - postgres
 
+
 volumes:
   kafka_data: {}
   pg_data: {}
diff --git a/fraud_service/app/main.py b/fraud_service/app/main.py
new file mode 100644
index 0000000..43a203c
--- /dev/null
+++ b/fraud_service/app/main.py
@@ -0,0 +1,125 @@
+import json
+import os
+import threading
+import time
+from contextlib import contextmanager
+from typing import Any, Dict, Optional, List
+
+import mlflow
+import numpy as np
+from fastapi import FastAPI
+from kafka import KafkaConsumer
+from pydantic import BaseModel, Field
+from sqlalchemy import JSON, Column, DateTime, Integer, MetaData, String, Table, create_engine, text
+from sqlalchemy.dialects.postgresql import DOUBLE_PRECISION
+from sqlalchemy.sql import func
+
+
+
+KAFKA_BOOTSTRAP_SERVERS=os.getenv("KAFKA_BOOTSTRAP_SERVERS","kafka:9092")
+KAFKA_TOPIC=os.getenv("KAFKA_TOPIC","transactions")
+POSTGRES_URI=os.getenv(
+    "POSTGRES_URI",
+    "postgresql+psycopg2://admin:admin@postgres:5432/frauddb",
+)
+
+MLFLOW_TRACKING_URI = os.getenv("MLFLOW_TRACKING_URI", "http://mlflow:5000")
+MLFLOW_MODEL_NAME = os.getenv("MLFLOW_MODEL_NAME", "fraud_detector")
+MLFLOW_MODEL_STAGE = os.getenv("MLFLOW_MODEL_STAGE", "Production")
+
+app = FastAPI(title="Fraud Detection Inference Service")
+
+
+# Database setup
+engine=create_engine(POSTGRES_URI,pool_pre_ping=True)
+metadata=MetaData()
+predictions_table=Table(
+    "predictions",
+    metadata,
+    Column("id",Integer,primary_key=True,utoincrement=True),
+    Column("event_time", DateTime(timezone=True), server_default=func.now()),
+    Column("transaction_id", String(64), nullable=False),
+    Column("amount", DOUBLE_PRECISION, nullable=False),
+    Column("features", JSON, nullable=False),
+    Column("prediction", Integer, nullable=False),
+    Column("proba", DOUBLE_PRECISION),
+    Column("model_version", String(255)),
+    Column("raw_payload", JSON),
+    schema="public",
+    extend_existing=True,
+    )
+
+
+def ensure_table():
+    with engine.begin() as conn:
+        conn.execute(text("CREATE SCHEMA IF NOT EXISTS public"))
+        metadata.create_all(conn)
+
+
+class Transaction(BaseModel):
+    transaction_id: str
+    amount: float
+    features: Dict[str, Any] = Field(default_factory=dict)
+
+FEATURE_ORDER=List[str]= ["amount", "num_items", "merchant_risk", "hour"]
+
+
+class ModelWrapper:
+    def __init__(self)->None:
+        self._model: Optional[Any]=None
+        self._model_version: Optional[Any]=None
+        mlflow.set_tracking_uri(MLFLOW_TRACKING_URI)
+
+    def load(self)->None:
+        try:
+            client=mlflow.tracking.MlflowClient()
+            if MLFLOW_MODEL_STAGE:
+                mv=client.get_latest_versions(MLFLOW_MODEL_NAME,stages=[MLFLOW_MODEL_STAGE])
+                if not mv:
+                    self._model=None
+                    self._model_version=None
+                    return
+                run_id=mv[0].version
+                model_uri=f"runs:/{run_id}/model"
+            else:
+                versions=client.get_latest_versions(MLFLOW_MODEL_NAME)
+                if not versions:
+                    self._model=None
+                    self._model_version=None
+                    return
+                run_id=versions[0].run_id
+                self._model_version=versions[0].version
+                model_uri=f"runs:/{run_id}/model"
+            self._model=mlflow.pyfunc.load_model(model_uri)
+        except Exception:
+            # Model may not exist yet
+            self._model=None
+            self._model_version=None
+
+    def predict_proba(self,features:Dict[str,Any])->float:
+        if self._model is None:
+            print("No Model Yet")
+            # Fallback: simple heuristic based on amount
+            amount=float(features.get('amount',0.0))
+            return 1.0 if amount>1000 else 0.05
+
+        # Project Features in the training order
+        x=np.array([[float(features.get(k,0.0)) for k in FEATURE_ORDER]],dtype=float)
+        try:
+            # pyfunc models don't always expose predict_proba; call predict
+            y=self._model.predict(x)
+
+            # If shape is (n,2), treat as proba for [0,1]
+            if isinstance(y,(list,tuple,np.ndarray)):
+                arr=np.array(y)
+                if arr.ndim==2 and arr.shape[1]>=2:
+                    return float(arr[0][1])
+                return float(arr[0])
+            return float(y)
+        except Exception:
+            amount=float(features.get('amount',0.0))
+            return 1.0 if amount>1000 else 0.05
+    @property
+    def version(self)-> Optional[str]:
+        return self._model_version
+