Building a CTR Optimization Pipeline from Scratch

import json
import os
import joblib
import pandas as pd
import csv
import random
from datetime import datetime
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from colorama import Fore, Style, init

init(autoreset=True)

app = FastAPI(title="Uber Eats API (A/B Test Mode)")

# --- CONFIGURATION ---
MODELS = {
    "v1": joblib.load("model_v1.joblib"), # Champion
    "v2": joblib.load("model_v2.joblib")  # Challenger
}
LOG_FILE = "ab_test_log.csv"

# Columns expected by models
EXPECTED_COLS = [
    'user_budget', 'is_subscribed', 'price_tier', 'rating', 'distance_km',
    'fav_cuisine_Burgers', 'fav_cuisine_Healthy', 'fav_cuisine_Indian', 
    'fav_cuisine_Pizza', 'fav_cuisine_Sushi',
    'cuisine_Burgers', 'cuisine_Healthy', 'cuisine_Indian', 
    'cuisine_Pizza', 'cuisine_Sushi'
]

# Create Log File
if not os.path.exists(LOG_FILE):
    with open(LOG_FILE, mode='w', newline='') as f:
        writer = csv.writer(f)
        writer.writerow(["timestamp", "model_used", "confidence", "recommended"])

class InteractionRequest(BaseModel):
    user_id: int = 0       # Added user_id for consistent routing
    user_budget: int
    fav_cuisine: str
    is_subscribed: int
    cuisine: str
    price_tier: int
    rating: float
    distance_km: float

@app.post("/predict")
def predict(request: InteractionRequest):
    
    # 1. ROUTING LOGIC (The A/B Split)
    # If user_id is even -> V1. If odd -> V2.
    # This ensures the same user always gets the same model experience.
    if request.user_id % 2 == 0:
        selected_version = "v1"
        model = MODELS["v1"]
        color = Fore.GREEN
    else:
        selected_version = "v2"
        model = MODELS["v2"]
        color = Fore.CYAN

    # 2. Preprocess (Standard)
    req_dict = request.dict()
    df_processed = pd.DataFrame(0, index=[0], columns=EXPECTED_COLS)
    
    df_processed['user_budget'] = req_dict['user_budget']
    df_processed['is_subscribed'] = req_dict['is_subscribed']
    df_processed['price_tier'] = req_dict['price_tier']
    df_processed['rating'] = req_dict['rating']
    df_processed['distance_km'] = req_dict['distance_km']
    
    fav_col = f"fav_cuisine_{req_dict['fav_cuisine']}"
    rest_col = f"cuisine_{req_dict['cuisine']}"
    if fav_col in df_processed.columns: df_processed[fav_col] = 1
    if rest_col in df_processed.columns: df_processed[rest_col] = 1

    # 3. Predict (Using ONLY the selected model)
    prob_numpy = model.predict_proba(df_processed)[0][1]
    prob = float(prob_numpy)
    rec = prob >= 0.5

    # 4. Log the Result (Crucial for Analysis)
    with open(LOG_FILE, mode='a', newline='') as f:
        writer = csv.writer(f)
        writer.writerow([
            datetime.now().strftime("%H:%M:%S"),
            selected_version,
            round(prob, 4),
            int(rec)
        ])

    # 5. Console Output
    rec_str = "REC" if rec else "SKIP"
    print(f"User {request.user_id:<3} -> {color}Served by {selected_version.upper()}{Style.RESET_ALL} | Conf: {prob:.2f} ({rec_str})")

    return {
        "recommended": bool(rec),
        "confidence": prob,
        "model_served": selected_version
    }