diff --git a/klausur-service/backend/ocr_pipeline_api.py b/klausur-service/backend/ocr_pipeline_api.py
index 82e745b..7265acb 100644
--- a/klausur-service/backend/ocr_pipeline_api.py
+++ b/klausur-service/backend/ocr_pipeline_api.py
@@ -1416,9 +1416,11 @@ async def _word_batch_stream_generator(
     # 2. Send preparing event (keepalive for proxy)
     yield f"data: {json.dumps({'type': 'preparing', 'message': 'Cell-First OCR laeuft parallel...'})}\n\n"
 
-    # 3. Run batch OCR in thread pool (CPU-bound, don't block event loop)
+    # 3. Run batch OCR in thread pool with periodic keepalive events.
+    # The OCR takes 30-60s and proxy servers (Nginx) may drop idle SSE
+    # connections after 30-60s. Send keepalive every 5s to prevent this.
     loop = asyncio.get_event_loop()
-    cells, columns_meta = await loop.run_in_executor(
+    ocr_future = loop.run_in_executor(
         None,
         lambda: build_cell_grid_v2(
             ocr_img, col_regions, row_geoms, img_w, img_h,
@@ -1426,6 +1428,25 @@ async def _word_batch_stream_generator(
         ),
     )
 
+    # Send keepalive events every 5 seconds while OCR runs
+    keepalive_count = 0
+    while not ocr_future.done():
+        try:
+            cells, columns_meta = await asyncio.wait_for(
+                asyncio.shield(ocr_future), timeout=5.0,
+            )
+            break  # OCR finished
+        except asyncio.TimeoutError:
+            keepalive_count += 1
+            elapsed = int(time.time() - t0)
+            yield f"data: {json.dumps({'type': 'keepalive', 'elapsed': elapsed, 'message': f'OCR laeuft... ({elapsed}s)'})}\n\n"
+            if await request.is_disconnected():
+                logger.info(f"SSE batch: client disconnected during OCR for {session_id}")
+                ocr_future.cancel()
+                return
+    else:
+        cells, columns_meta = ocr_future.result()
+
     if await request.is_disconnected():
         logger.info(f"SSE batch: client disconnected after OCR for {session_id}")
         return