John Patrick TeruelandCopilot
4cefdb9083
- Added consumer/Dockerfile with all dependencies - Updated consumer pipeline.py to use environment variables - Updated features.py and decision.py to support env var configuration - Updated docker-compose.yml to include consumer service - Consumer now runs automatically with 'docker compose up' - Proper Kafka and Redis networking for Docker - Environment variables for ML_MODELS_DIR, KAFKA_BROKER, REDIS_HOST/PORT Now entire PayGuard system runs with: docker compose up -d Co-authored-by: Copilot <[email protected]>
94 lines
3.2 KiB
Python
94 lines
3.2 KiB
Python
"""
|
|
Model Scoring Modules
|
|
|
|
Responsibilities:
|
|
- Load pre-trained models
|
|
- Run Layer 1 (fast) and Layer 2 (deep) scoring
|
|
- Return fraud probabilities
|
|
|
|
Academic Point:
|
|
Dual-layer architecture allows high-throughput systems to make most decisions
|
|
(>95% of transactions) with a fast, lightweight model, while complex edge cases
|
|
are delegated to a more expensive deep model. This optimizes latency vs. accuracy.
|
|
"""
|
|
|
|
import pickle
|
|
import logging
|
|
import os
|
|
from typing import Tuple
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
# Import model classes - they may be at /app/ml_models or /models in different contexts
|
|
try:
|
|
# Docker context
|
|
from ml_models import Layer1MockModel, Layer2MockModel
|
|
except ImportError:
|
|
try:
|
|
# Local context (when running from host)
|
|
from models import Layer1MockModel, Layer2MockModel
|
|
except ImportError:
|
|
pass
|
|
|
|
class ModelScorer:
|
|
"""Loads and runs fraud detection models"""
|
|
|
|
def __init__(self, layer1_path: str = None, layer2_path: str = None):
|
|
# Auto-detect paths based on context
|
|
if layer1_path is None:
|
|
models_dir = os.getenv("ML_MODELS_DIR", "models")
|
|
layer1_path = os.path.join(models_dir, "layer1.pkl")
|
|
|
|
if layer2_path is None:
|
|
models_dir = os.getenv("ML_MODELS_DIR", "models")
|
|
layer2_path = os.path.join(models_dir, "layer2.pkl")
|
|
|
|
# Load Layer 1 model
|
|
with open(layer1_path, 'rb') as f:
|
|
self.layer1_model = pickle.load(f)
|
|
logger.info(f"[MODEL LOADER] ✓ Layer 1 model loaded from {layer1_path}")
|
|
|
|
# Load Layer 2 model
|
|
with open(layer2_path, 'rb') as f:
|
|
self.layer2_model = pickle.load(f)
|
|
logger.info(f"[MODEL LOADER] ✓ Layer 2 model loaded from {layer2_path}")
|
|
|
|
def score_layer1(self, features) -> float:
|
|
"""
|
|
Layer 1: Fast, shallow model for initial risk assessment.
|
|
|
|
Decision threshold: 0.4
|
|
- score < 0.4 → Low risk (approve immediately)
|
|
- score >= 0.4 → Suspicious (escalate to Layer 2)
|
|
"""
|
|
probs = self.layer1_model.predict_proba(features)
|
|
fraud_probability = probs[0][1] # Index 1 is fraud class
|
|
|
|
logger.info(
|
|
f"[LAYER 1 SCORING] Fraud probability: {fraud_probability:.4f} "
|
|
f"{'→ LOW RISK, fast-track approval' if fraud_probability < 0.4 else '→ SUSPICIOUS, escalate to Layer 2'}"
|
|
)
|
|
|
|
return fraud_probability
|
|
|
|
def score_layer2(self, features) -> float:
|
|
"""
|
|
Layer 2: Deep, complex model for suspicious transactions.
|
|
|
|
Decision threshold: 0.7
|
|
- score <= 0.7 → Approve
|
|
- score > 0.7 → Block
|
|
|
|
Only called if Layer 1 score >= 0.4 (suspicious).
|
|
"""
|
|
probs = self.layer2_model.predict_proba(features)
|
|
fraud_probability = probs[0][1] # Index 1 is fraud class
|
|
|
|
logger.info(
|
|
f"[LAYER 2 SCORING] Fraud probability: {fraud_probability:.4f} "
|
|
f"{'→ BLOCK (high confidence in fraud)' if fraud_probability > 0.7 else '→ APPROVE (acceptable risk)'}"
|
|
)
|
|
|
|
return fraud_probability
|
|
|