diff --git a/Dockerfile b/Dockerfile
index a8b6358..4228c2c 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -43,9 +43,22 @@ COPY backend ./backend
 COPY locales ./locales
 COPY package.json ./
 
-# Install backend deps against the final pyproject.toml/uv.lock
+# Install backend deps against the final pyproject.toml/uv.lock, then strip the
+# NVIDIA CUDA runtime libraries that torch drags in (via camel-oasis ->
+# sentence-transformers). The server has no GPU and all LLM calls go through an
+# API, so CUDA libs are pure bloat (many GB). torch itself stays (CPU runtime)
+# but without the nvidia-* packages the image is dramatically smaller. We then
+# verify the stripped env still imports torch/sentence-transformers and the app
+# so the build fails loudly instead of a silent runtime break.
 RUN cd backend && uv sync --frozen --no-dev \
-  && uv run --frozen python -c "import psycopg" || (echo "FATAL: psycopg not installed — package sync missing the Postgres driver" >&2 && exit 1)
+  && uv pip uninstall \
+       nvidia-cublas-cu12 nvidia-cuda-cupti-cu12 nvidia-cuda-nvrtc-cu12 \
+       nvidia-cuda-runtime-cu12 nvidia-cudnn-cu12 nvidia-cufft-cu12 \
+       nvidia-cufile-cu12 nvidia-curand-cu12 nvidia-cusolver-cu12 \
+       nvidia-cusparse-cu12 nvidia-cusparselt-cu12 nvidia-nccl-cu12 \
+       nvidia-nvjitlink-cu12 nvidia-nvshmem-cu12 nvidia-nvtx-cu12 triton \
+  && uv run --frozen python -c "import psycopg, torch; import sentence_transformers; from app import create_app; print('deps-after-strip-ok')" \
+    || (echo "FATAL: dependency import failed after CUDA strip (torch/sentence-transformers/app). Revisit the nvidia-* uninstall list." >&2 && exit 1)
 
 # Make the migration-runner entrypoint executable
 RUN chmod +x /app/backend/docker_entrypoint.sh