diff --git a/backend/app/services/ml/tagger.py b/backend/app/services/ml/tagger.py
index 0ee33eb..40ad21c 100644
--- a/backend/app/services/ml/tagger.py
+++ b/backend/app/services/ml/tagger.py
@@ -15,9 +15,13 @@ from dataclasses import dataclass
 from pathlib import Path
 
 import numpy as np
-import onnxruntime as ort
 from PIL import Image, ImageFile
 
+# onnxruntime lives in requirements-ml.txt only — it is NOT installed in the
+# lean web image or in CI. Imported lazily inside Tagger.load() so this module
+# imports fine without it (the suggestion service imports SURFACED_CATEGORIES
+# from here in the web container, and CI collects the pure-logic tests).
+
 # Tolerate minutely-truncated source images (same rationale as IR's wd14.py:
 # a few missing bytes at the JPEG EOI shouldn't block tagging the whole image).
 ImageFile.LOAD_TRUNCATED_IMAGES = True
@@ -43,7 +47,7 @@ class TagPrediction:
 class Tagger:
     def __init__(self, model_dir: Path | None = None):
         self._model_dir = model_dir or _MODEL_DIR
-        self._session: ort.InferenceSession | None = None
+        self._session = None  # onnxruntime.InferenceSession once load()ed
         self._tag_meta: list[dict] | None = None
         self._input_name: str | None = None
         self._output_name: str | None = None
@@ -73,6 +77,11 @@ class Tagger:
                     {"name": row["name"], "category": row["category"]}
                 )
 
+        # Lazy import — kept after the file-existence checks so the
+        # missing-model RuntimeError still fires first in environments
+        # without onnxruntime (CI / lean web image).
+        import onnxruntime as ort
+
         session = ort.InferenceSession(
             str(model_path), providers=["CPUExecutionProvider"]
         )