diff --git a/htx/htx.py b/htx/htx.py
index fe73a95..8dcb096 100644
--- a/htx/htx.py
+++ b/htx/htx.py
@@ -1,7 +1,9 @@
 import json
 import multiprocessing as mp
 import logging
+import threading
 
+from queue import Queue
 from flask import Flask, request, jsonify
 
 from htx.model_manager import ModelManager
@@ -19,9 +21,9 @@ def init_model_server(**kwargs):
 
 @_server.before_first_request
 def launch_train_loop():
-    train_process = mp.Process(
+    train_process = threading.Thread(
         target=_model_manager.train_loop,
-        args=(_model_manager.queue, _model_manager.train_script)
+        args=(_model_manager.queue, _model_manager.train_script, _model_manager.redis, )
     )
     train_process.start()
 
diff --git a/htx/model_manager.py b/htx/model_manager.py
index 9555461..c8ffced 100644
--- a/htx/model_manager.py
+++ b/htx/model_manager.py
@@ -1,10 +1,10 @@
 import os
-import multiprocessing as mp
 import logging
 import json
 import attr
 import io
 import shutil
+import queue
 
 from operator import attrgetter
 from redis import Redis
@@ -16,6 +16,18 @@
 logger = logging.getLogger(__name__)
 
 
+@attr.s
+class ModelWrapper(object):
+    model = attr.ib()
+    version = attr.ib()
+
+    def predict(self, *args, **kwargs):
+        return self.model.predict(*args, **kwargs)
+
+    def get_data_item(self, *args, **kwargs):
+        return self.model.get_data_item(*args, **kwargs)
+
+
 class QueuedItem(object):
 
     def __init__(self, project):
@@ -54,7 +66,7 @@ class ModelManager(object):
 
     _MODEL_LIST_FILE = 'model_list.txt'
     _DEFAULT_MODEL_VERSION = 'model'
-    queue = mp.Queue()
+    queue = queue.Queue()
 
     def __init__(
         self,
@@ -64,6 +76,7 @@ def __init__(
         data_dir='~/.heartex/data',
         min_examples_for_train=1,
         retrain_after_num_examples=1,
+        redis=None,
         redis_host='localhost',
         redis_port=6379,
         redis_queue='default',
@@ -87,11 +100,13 @@ def __init__(
         self.model_list_file = os.path.join(self.model_dir, self._MODEL_LIST_FILE)
 
         self._current_model = {}
-        self._redis = Redis(host=redis_host, port=redis_port)
+        if redis is not None:
+            self._redis = redis
+        else:
+            self._redis = Redis(host=redis_host, port=redis_port)
 
     def _get_latest_finished_train_job(self, project):
-        queue = Queue(name=self.redis_queue, connection=self._redis)
-        registry = FinishedJobRegistry(queue.name, queue.connection)
+        registry = FinishedJobRegistry(self.redis_queue, self._redis)
         if registry.count == 0:
             logger.info('Train job registry is empty.')
             return None
@@ -140,7 +155,7 @@ def setup(self, project, schema):
                 logger.error(f'Found resources {resources}, but model is not loaded for project {project}. '
                              f'Consequent API calls (e.g. predict) will fail.')
                 return None
-        self._current_model[project] = model
+        self._current_model[project] = ModelWrapper(model, model_version)
         self._stash_resources(project, resources)
         logger.info(f'Model {model_version} successfully loaded for project {project}.')
         return model_version
@@ -283,8 +298,7 @@ def _flush_all(self, project, redis, reqis_queue):
             # TODO: do we need the locks here?
             shutil.rmtree(project_data_dir)
 
-    def train_loop(self, data_queue, train_script):
-        redis = Redis(host=self.redis_host, port=self.redis_port)
+    def train_loop(self, data_queue, train_script, redis):
         redis_queue = Queue(name=self.redis_queue, connection=redis)
         logger.info(f'Train loop starts: PID={os.getpid()}, Redis connection: {redis}, queue: {redis_queue}')
         for queued_items, in iter(data_queue.get, None):