diff --git a/storey/sources.py b/storey/sources.py
index 95aba149..c215192b 100644
--- a/storey/sources.py
+++ b/storey/sources.py
@@ -1,4 +1,5 @@
 import asyncio
+import copy
 import csv
 import math
 import queue
@@ -510,6 +511,7 @@ def __init__(self, paths: Union[List[str], str], header: bool = False, build_dic
     def _init(self):
         self._event_buffer = queue.Queue(1024)
         self._types = []
+        self._none_columns = set()
 
     def _infer_type(self, value):
         lowercase = value.lower()
@@ -534,6 +536,9 @@ def _infer_type(self, value):
         except ValueError:
             pass
 
+        if value == '':
+            return 'n'
+
         return 's'
 
     def _parse_field(self, field, index):
@@ -557,6 +562,8 @@ def _parse_field(self, field, index):
             if field == '':
                 return None
             return self._datetime_from_timestamp(field)
+        if typ == 'n':
+            return None
         raise TypeError(f'Unknown type: {typ}')
 
     def _datetime_from_timestamp(self, timestamp):
@@ -583,8 +590,17 @@ def _blocking_io_loop(self):
                         parsed_line = next(csv.reader([line]))
                         if self._type_inference:
                             if not self._types:
-                                for field in parsed_line:
-                                    self._types.append(self._infer_type(field))
+                                for index, field in enumerate(parsed_line):
+                                    type_field = self._infer_type(field)
+                                    self._types.append(type_field)
+                                    if type_field == 'n':
+                                        self._none_columns.add(index)
+                            else:
+                                for index in copy.copy(self._none_columns):
+                                    type_field = self._infer_type(parsed_line[index])
+                                    if type_field != 'n':
+                                        self._types[index] = type_field
+                                        self._none_columns.remove(index)
                             for i in range(len(parsed_line)):
                                 parsed_line[i] = self._parse_field(parsed_line[i], i)
                         element = parsed_line
diff --git a/tests/test_flow.py b/tests/test_flow.py
index 19f84f35..f867bb45 100644
--- a/tests/test_flow.py
+++ b/tests/test_flow.py
@@ -2554,3 +2554,25 @@ def test_custom_string_time_input():
     result = awaitable_result.await_result()
     controller.terminate()
     assert result.time == datetime(2021, 5, 9, 14, 5, 27, tzinfo=pytz.utc)
+
+
+def test_csv_none_value_first_row(tmpdir):
+    out_file_par = f'{tmpdir}/test_csv_none_value_first_row_{uuid.uuid4().hex}.parquet'
+    out_file_csv = f'{tmpdir}/test_csv_none_value_first_row_{uuid.uuid4().hex}.csv'
+
+    columns = ['first_name', "bid", "bool", "time"]
+    data = pd.DataFrame([['katya', None, None, None], ['dina', 45.7, True, datetime(2021, 4, 21, 15, 56, 53, 385444)]],
+                        columns=columns)
+    data.to_csv(out_file_csv)
+
+    controller = build_flow([
+        CSVSource(out_file_csv, header=True, key_field='first_name', build_dict=True),
+        ParquetTarget(out_file_par)
+    ]).run()
+
+    controller.await_termination()
+    read_back_df = pd.read_parquet(out_file_par)
+
+    for c in columns:
+        assert read_back_df.dtypes.to_dict()[c] == data.dtypes.to_dict()[c]
+