diff --git a/lab/serving/__init__.py b/lab/serving/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/lab/serving/api.py b/lab/serving/api.py
new file mode 100644
index 0000000..1da2503
--- /dev/null
+++ b/lab/serving/api.py
@@ -0,0 +1,99 @@
+"""API REST de prediction de consommation electrique (FastAPI).
+
+Le service orchestre la chaine de prediction :
+  1. recevoir la requete (client_id, date) ;
+  2. recuperer les features du client (feature store simule) ;
+  3. charger le modele promu (Model Registry, au demarrage) ;
+  4. calculer la prediction et repondre en JSON.
+
+Lancement : uvicorn lab.serving.api:app --host 0.0.0.0 --port 8000
+Swagger UI : /docs
+"""
+
+import logging
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI, HTTPException
+
+from . import features, registry
+from .schemas import (
+    BatchPredictionRequest,
+    BatchPredictionResponse,
+    PredictionRequest,
+    PredictionResponse,
+)
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+# Modele charge une fois au demarrage (couteux) et reutilise a chaque requete.
+_state: dict[str, registry.LoadedModel] = {}
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    _state["model"] = registry.load_champion()
+    yield
+    _state.clear()
+
+
+app = FastAPI(
+    title="Electricity Consumption Prediction API",
+    description="Expose le modele promu (MLflow Model Registry) via une API REST.",
+    version="1.0.0",
+    lifespan=lifespan,
+)
+
+
+def _get_model() -> registry.LoadedModel:
+    model = _state.get("model")
+    if model is None:  # modele indisponible au demarrage
+        raise HTTPException(status_code=503, detail="Modele non charge.")
+    return model
+
+
+def _predict(client_id: str, model: registry.LoadedModel) -> PredictionResponse:
+    feats = features.get_features(client_id)  # peut lever UnknownClientError
+    prediction = model.predict_one(feats)
+    return PredictionResponse(
+        client_id=client_id,
+        prediction_kwh=prediction,
+        model_name=model.name,
+        model_version=model.version,
+    )
+
+
+@app.get("/health", summary="Verification de l'etat du service")
+def health() -> dict[str, str]:
+    """Endpoint de sante : renvoie 200 si le service repond."""
+    return {"status": "ok"}
+
+
+@app.post("/predict", response_model=PredictionResponse, summary="Prediction unitaire")
+def predict(request: PredictionRequest) -> PredictionResponse:
+    model = _get_model()
+    try:
+        return _predict(request.client_id, model)
+    except features.UnknownClientError:
+        # Partie 3 : client inconnu -> erreur cliente, pas un plantage du service.
+        raise HTTPException(
+            status_code=404,
+            detail=f"Client inconnu : aucune feature pour '{request.client_id}'.",
+        )
+
+
+@app.post(
+    "/predict/batch",
+    response_model=BatchPredictionResponse,
+    summary="Prediction en batch (plusieurs clients)",
+)
+def predict_batch(request: BatchPredictionRequest) -> BatchPredictionResponse:
+    model = _get_model()
+    predictions: list[PredictionResponse] = []
+    unknown: list[str] = []
+    for client_id in request.client_ids:
+        try:
+            predictions.append(_predict(client_id, model))
+        except features.UnknownClientError:
+            unknown.append(client_id)
+    return BatchPredictionResponse(predictions=predictions, unknown_client_ids=unknown)
diff --git a/lab/serving/features.py b/lab/serving/features.py
new file mode 100644
index 0000000..1c85672
--- /dev/null
+++ b/lab/serving/features.py
@@ -0,0 +1,66 @@
+"""Recuperation des features de prediction.
+
+Dans un systeme reel, ces valeurs proviendraient d'un feature store / d'une base
+alimentee par le pipeline de calcul de features (lags, moyennes glissantes) sur
+l'historique de consommation. On separe volontairement cette etape du calcul de la
+prediction : la source des features peut changer sans toucher au modele.
+
+Ici on SIMULE cette recuperation par un simple dictionnaire Python, comme demande
+par l'enonce. Les valeurs sont des observations reelles (derniere ligne connue de
+quelques clients dans data/test.parquet).
+"""
+
+from .. import constants
+
+
+class UnknownClientError(KeyError):
+    """Aucune feature disponible pour ce client (identifiant inconnu)."""
+
+
+# feature store simule : client_id -> {feature: valeur}
+FEATURE_STORE: dict[str, dict[str, float]] = {
+    "MT_124": {
+        "lag_1d": 107.656,
+        "lag_7d": 25.120,
+        "lag_30d": 70.574,
+        "lag_365d": 25.120,
+        "rolling_mean_7d": 65.870,
+        "rolling_mean_30d": 71.310,
+    },
+    "MT_156": {
+        "lag_1d": 13.149,
+        "lag_7d": 13.929,
+        "lag_30d": 21.577,
+        "lag_365d": 8.935,
+        "rolling_mean_7d": 16.648,
+        "rolling_mean_30d": 19.720,
+    },
+    "MT_158": {
+        "lag_1d": 30.739,
+        "lag_7d": 16.608,
+        "lag_30d": 34.094,
+        "lag_365d": 6.574,
+        "rolling_mean_7d": 19.067,
+        "rolling_mean_30d": 21.486,
+    },
+    "MT_159": {
+        "lag_1d": 23.305,
+        "lag_7d": 24.741,
+        "lag_30d": 21.386,
+        "lag_365d": 5.333,
+        "rolling_mean_7d": 11.707,
+        "rolling_mean_30d": 13.619,
+    },
+}
+
+
+def get_features(client_id: str) -> dict[str, float]:
+    """Renvoyer les features du client, ordonnees comme a l'entrainement du modele.
+
+    Leve UnknownClientError si le client est inconnu.
+    """
+    if client_id not in FEATURE_STORE:
+        raise UnknownClientError(client_id)
+    raw = FEATURE_STORE[client_id]
+    # On respecte l'ordre des colonnes attendu par le modele (SERVING_FEATURES).
+    return {feature: raw[feature] for feature in constants.SERVING_FEATURES}
diff --git a/lab/serving/registry.py b/lab/serving/registry.py
new file mode 100644
index 0000000..71e91b1
--- /dev/null
+++ b/lab/serving/registry.py
@@ -0,0 +1,43 @@
+"""Chargement du modele promu depuis le MLflow Model Registry (par alias)."""
+
+import logging
+from dataclasses import dataclass
+
+import mlflow
+import pandas as pd
+
+from .. import constants
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class LoadedModel:
+    """Modele charge + metadonnees de version, garde en memoire par l'API."""
+
+    model: mlflow.pyfunc.PyFuncModel
+    name: str
+    version: str
+
+    def predict_one(self, features: dict[str, float]) -> float:
+        """Prediction pour un jeu de features (1 ligne)."""
+        frame = pd.DataFrame([features], columns=constants.SERVING_FEATURES)
+        return float(self.model.predict(frame)[0])
+
+
+def load_champion() -> LoadedModel:
+    """Charger la version pointee par l'alias `MODEL_ALIAS` du modele enregistre.
+
+    On identifie le modele par `models:/<nom>@<alias>` : le meme code sert n'importe
+    quelle version promue, sans redeploiement, en changeant seulement l'alias cote MLflow.
+    """
+    name = constants.REGISTERED_MODEL_NAME
+    alias = constants.MODEL_ALIAS
+    uri = f"models:/{name}@{alias}"
+    logger.info(f"Chargement du modele {uri}")
+
+    client = mlflow.MlflowClient()
+    version = client.get_model_version_by_alias(name=name, alias=alias)
+    model = mlflow.pyfunc.load_model(uri)
+    logger.info(f"Modele charge : {name} v{version.version}")
+    return LoadedModel(model=model, name=name, version=str(version.version))
diff --git a/lab/serving/schemas.py b/lab/serving/schemas.py
new file mode 100644
index 0000000..e53bc5d
--- /dev/null
+++ b/lab/serving/schemas.py
@@ -0,0 +1,44 @@
+"""Schemas Pydantic du service de prediction (contrat d'entree/sortie de l'API)."""
+
+import datetime
+
+from pydantic import BaseModel, Field
+
+
+class PredictionRequest(BaseModel):
+    """Ce que le client fournit : QUI et QUAND, pas les features (calculees cote serveur)."""
+
+    client_id: str = Field(
+        ...,
+        description="Identifiant du client (ex. 'MT_124').",
+        examples=["MT_124"],
+    )
+    date: datetime.date | None = Field(
+        default=None,
+        description="Date de la prediction (parametre de requete). Optionnelle.",
+        examples=["2015-01-01"],
+    )
+
+
+class BatchPredictionRequest(BaseModel):
+    """Prediction pour plusieurs clients en une seule requete (inference batch)."""
+
+    client_ids: list[str] = Field(
+        ...,
+        description="Liste d'identifiants clients.",
+        examples=[["MT_124", "MT_156", "MT_158"]],
+    )
+    date: datetime.date | None = None
+
+
+class PredictionResponse(BaseModel):
+    client_id: str
+    prediction_kwh: float = Field(description="Consommation predite (kWh).")
+    model_name: str
+    model_version: str
+
+
+class BatchPredictionResponse(BaseModel):
+    predictions: list[PredictionResponse]
+    # Clients ignores (features introuvables) : on ne fait pas echouer tout le lot.
+    unknown_client_ids: list[str] = Field(default_factory=list)
diff --git a/serve.sh b/serve.sh
new file mode 100755
index 0000000..c5abaa0
--- /dev/null
+++ b/serve.sh
@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+# Lance l'API de prediction (TP03). A executer depuis la racine du depot (/home/user/tp).
+# Charge .env (MLflow + creds S3 pour recuperer le modele promu depuis le Registry).
+set -euo pipefail
+cd "$(dirname "$0")"
+set -a
+# shellcheck disable=SC1091
+source .env
+set +a
+exec /opt/venvs/mlops/bin/uvicorn lab.serving.api:app --host 0.0.0.0 --port 8000