diff --git a/.env b/.env
index b3e66921..d7ce1b3c 100644
--- a/.env
+++ b/.env
@@ -46,10 +46,10 @@ GUNICORN_WORKERS=1
 # It works because we added the special `host.docker.internal:host-gateway`
 # host in dev.yml for all services
 # Triton is the ML inference server used at Open Food Facts
-TRITON_URI=host.docker.internal:5004
+TRITON_URI=host.docker.internal:5504
 
 # By default, don't enable ML predictions, as we don't necessarily have a Triton
 # server running.
 # During local development, to enable ML predictions, set this to True and make sure
-# you have Triton running on port 5004.
+# you have Triton running on port 5504.
 ENABLE_ML_PREDICTIONS=False
diff --git a/config/settings.py b/config/settings.py
index a4ef7da1..80a46a86 100644
--- a/config/settings.py
+++ b/config/settings.py
@@ -293,5 +293,5 @@
 # Triton Inference Server (ML)
 # ------------------------------------------------------------------------------
 
-TRITON_URI = os.getenv("TRITON_URI", "localhost:5004")
+TRITON_URI = os.getenv("TRITON_URI", "localhost:5504")
 ENABLE_ML_PREDICTIONS = os.getenv("ENABLE_ML_PREDICTIONS") == "True"
diff --git a/open_prices/proofs/management/commands/run_ml_model.py b/open_prices/proofs/management/commands/run_ml_model.py
new file mode 100644
index 00000000..6267ee87
--- /dev/null
+++ b/open_prices/proofs/management/commands/run_ml_model.py
@@ -0,0 +1,53 @@
+import argparse
+
+import tqdm
+from django.core.management.base import BaseCommand
+
+from open_prices.proofs.ml.image_classifier import (
+    PROOF_CLASSIFICATION_MODEL_NAME,
+    run_and_save_proof_prediction,
+)
+from open_prices.proofs.models import Proof
+
+
+class Command(BaseCommand):
+    help = """Run ML models on images with proof predictions, and save the predictions
+    in DB."""
+    _allowed_types = ["proof_classification"]
+
+    def add_arguments(self, parser: argparse.ArgumentParser) -> None:
+        parser.add_argument(
+            "--limit", type=int, help="Limit the number of proofs to process."
+        )
+        parser.add_argument("type", type=str, help="Type of model to run.", nargs="+")
+
+    def handle(self, *args, **options) -> None:  # type: ignore
+        self.stdout.write(
+            "Running ML models on images without proof predictions for this model..."
+        )
+        limit = options["limit"]
+        types = options["type"]
+
+        if not all(t in self._allowed_types for t in types):
+            raise ValueError(
+                f"Invalid type(s) provided: {types}, allowed: {self._allowed_types}"
+            )
+
+        if "proof_classification" in types:
+            # Get proofs that don't have a proof prediction with
+            # model_name = PROOF_CLASSIFICATION_MODEL_NAME by performing an
+            # outer join on the Proof and Prediction tables.
+            proofs = (
+                Proof.objects.filter(predictions__model_name__isnull=True)
+                | Proof.objects.exclude(
+                    predictions__model_name=PROOF_CLASSIFICATION_MODEL_NAME
+                )
+            ).distinct()
+
+            if limit:
+                proofs = proofs[:limit]
+
+            for proof in tqdm.tqdm(proofs):
+                self.stdout.write(f"Processing proof {proof.id}...")
+                run_and_save_proof_prediction(proof.id)
+                self.stdout.write("Done.")