diff --git a/app.py b/app.py
index 86cb289..7935f30 100644
--- a/app.py
+++ b/app.py
@@ -79,16 +79,15 @@ class SeparatorWorker(core.Stack):
                  **kwargs):
         super().__init__(scope, id, **kwargs)
 
-        worker_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), 'worker-placeholder'))
+        worker_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), 'worker'))
 
         self.service = ecs_patterns.QueueProcessingFargateService(self, 'separator-service',
             cluster=cluster,
-            cpu=256,
-            memory_limit_mib=512,
+            cpu=2048,
+            memory_limit_mib=8192,
             image=ecs.ContainerImage.from_asset(directory=worker_dir),
             environment={
                 'TRACKS_TABLE_NAME': tracks_table.table_name,
-                'INPUT_BUCKET_NAME': input_bucket.bucket_name,
                 'OUTPUT_BUCKET_NAME': output_bucket.bucket_name
             })
 
diff --git a/worker/Dockerfile b/worker/Dockerfile
new file mode 100644
index 0000000..3c9c8ca
--- /dev/null
+++ b/worker/Dockerfile
@@ -0,0 +1,9 @@
+FROM researchdeezer/spleeter
+ENV PYTHONDONTWRITEBYTECODE=1
+
+WORKDIR /app
+COPY . /app
+RUN pip install --no-cache-dir -r requirements.txt
+RUN python -m prefetch_models
+
+ENTRYPOINT ["python", "separator.py"]
diff --git a/worker/conftest.py b/worker/conftest.py
new file mode 100644
index 0000000..e69de29
diff --git a/worker/prefetch_models.py b/worker/prefetch_models.py
new file mode 100644
index 0000000..7539c73
--- /dev/null
+++ b/worker/prefetch_models.py
@@ -0,0 +1,14 @@
+'''
+Pre-fetches the required model from github.
+'''
+
+import os
+from spleeter.model.provider import get_default_model_provider
+
+
+PREFETCH_MODELS = ['2stems']
+
+for model_name in PREFETCH_MODELS:
+    get_default_model_provider().download(
+        model_name,
+        os.path.join(os.getenv('MODEL_PATH'), model_name))
diff --git a/worker/pytest.ini b/worker/pytest.ini
new file mode 100644
index 0000000..9cc4abd
--- /dev/null
+++ b/worker/pytest.ini
@@ -0,0 +1,4 @@
+[pytest]
+filterwarnings =
+    ignore::FutureWarning
+    ignore::DeprecationWarning
diff --git a/worker/requirements.txt b/worker/requirements.txt
new file mode 100644
index 0000000..3f3536b
--- /dev/null
+++ b/worker/requirements.txt
@@ -0,0 +1,3 @@
+boto3
+pytest
+moto
diff --git a/worker/separator.py b/worker/separator.py
new file mode 100644
index 0000000..c2a565c
--- /dev/null
+++ b/worker/separator.py
@@ -0,0 +1,109 @@
+import os
+import logging
+import json
+import sys
+from collections import namedtuple
+from tempfile import NamedTemporaryFile, TemporaryDirectory
+from threading import Timer
+from urllib.parse import urlparse
+
+import boto3
+from spleeter.separator import Separator
+from spleeter.audio.adapter import get_audio_adapter
+
+
+SPLEETER_CONFIGURATION = os.getenv('SPLEETER_CONFIGURATION', 'spleeter:2stems')
+OUTPUT_CODEC = os.getenv('OUTPUT_CODEC', 'mp3')
+OUTPUT_BITRATE = os.getenv('OUTPUT_BITRATE', '128k')
+MAX_AUDIO_DURATION = float(os.getenv('MAX_AUDIO_DURATION', 600.))
+AUDIO_START_OFFSET = float(os.getenv('AUDIO_START_OFFSET', 0.))
+OUTPUT_FILENAME_FORMAT = os.getenv('OUTPUT_FILENAME_FORMAT', '{instrument}.{codec}')
+QUEUE_NAME = os.getenv('QUEUE_NAME')
+POLLING_INTERVAL = int(os.getenv('POLLING_INTERVAL', 5))
+OUTPUT_BUCKET_NAME = os.getenv('OUTPUT_BUCKET_NAME')
+TRACKS_TABLE_NAME = os.getenv('TRACKS_TABLE_NAME')
+USE_MULTICHANNEL_WIENER_FILTERING = False
+
+
+S3Entry = namedtuple('S3Entry', ['bucket_name', 'key'])
+
+
+def parse_s3_url(s3_url: str) -> S3Entry:
+    o = urlparse(s3_url)
+    return S3Entry(o.netloc, o.path.lstrip('/'))
+
+
+def fetch_separate_and_upload(input_s3_url, output_s3_url, s3=None):
+    if s3 is None:
+        s3 = boto3.resource("s3")
+
+    with NamedTemporaryFile() as input_file, TemporaryDirectory() as output_path:
+        input_object = s3.Object(**parse_s3_url(input_s3_url)._asdict())
+        input_object.download_file(input_file.name)
+
+        audio_adapter = get_audio_adapter(None)
+        separator = Separator(
+            SPLEETER_CONFIGURATION,
+            MWF=USE_MULTICHANNEL_WIENER_FILTERING)
+
+        separator.separate_to_file(
+            input_file.name,
+            output_path,
+            audio_adapter=audio_adapter,
+            offset=AUDIO_START_OFFSET,
+            duration=MAX_AUDIO_DURATION,
+            codec=OUTPUT_CODEC,
+            bitrate=OUTPUT_BITRATE,
+            filename_format=OUTPUT_FILENAME_FORMAT,
+            synchronous=True)
+
+        logging.info(f'Uploading output to: {output_s3_url}')
+        output_object = s3.Object(**parse_s3_url(output_s3_url)._asdict())
+
+        output_filename = os.path.join(output_path, f'accompaniment.{OUTPUT_CODEC}')
+        output_object.upload_file(output_filename)
+
+
+def poll_for_sqs_message(queue_name: str):
+    sqs = boto3.client('sqs')
+    queue_url = sqs.get_queue_url(QueueName=queue_name)['QueueUrl']
+    response = sqs.receive_message(QueueUrl=queue_url)
+
+    try:
+        messages = response['Messages']
+    except KeyError:
+        logging.info('No messages in the queue')
+        messages = []
+
+    for message in messages:
+        body = json.loads(message['Body'])
+        job_id = body['job_id']
+        output_s3_url = f"s3://{OUTPUT_BUCKET_NAME}/track_{job_id}.mp3"
+        logging.info(f'Start separating {job_id} -> {output_s3_url}')
+
+        try:
+            fetch_separate_and_upload(
+                input_s3_url=body['input_s3_url'],
+                output_s3_url=output_s3_url)
+
+            logging.info(f'Processing successful: {output_s3_url}')
+
+            dynamodb = boto3.client('dynamodb')
+            dynamodb.update_item(
+                TableName=TRACKS_TABLE_NAME,
+                Key={'id': {'S': job_id}},
+                AttributeUpdates={
+                    'status': {'Value': {'S': 'successful'}},
+                    'output_url': {'Value': {'S': output_s3_url}}
+                })
+        except:
+            logging.exception('Processing failed for some reason')
+        finally:
+            sqs.delete_message(QueueUrl=queue_url, ReceiptHandle=message['ReceiptHandle'])
+
+        Timer(5, poll_for_sqs_message, args=[queue_name]).start()
+
+
+if __name__ == '__main__':
+    logging.basicConfig(level=logging.INFO)
+    poll_for_sqs_message(QUEUE_NAME)
diff --git a/worker/tests/audio_example.mp3 b/worker/tests/audio_example.mp3
new file mode 100644
index 0000000..cc91788
Binary files /dev/null and b/worker/tests/audio_example.mp3 differ
diff --git a/worker/tests/test_separator.py b/worker/tests/test_separator.py
new file mode 100644
index 0000000..5b359a7
--- /dev/null
+++ b/worker/tests/test_separator.py
@@ -0,0 +1,39 @@
+import os
+
+import boto3
+import pytest
+from moto import mock_s3
+
+from separator import fetch_separate_and_upload, parse_s3_url
+
+
+TEST_FILENAME = 'audio_example.mp3'
+
+
+@pytest.fixture
+def mocked_cloud():
+    with mock_s3():
+        s3 = boto3.resource('s3', region_name='eu-east-1')
+        input_bucket = s3.create_bucket(Bucket='input')
+
+        test_file_local_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), TEST_FILENAME)
+        input_bucket.upload_file(Filename=test_file_local_path, Key=TEST_FILENAME)
+
+        yield {
+            's3': s3,
+            'input_bucket': input_bucket,
+            'output_bucket': s3.create_bucket(Bucket='output')
+        }
+
+
+def test_fetch_separate_and_upload(mocked_cloud):
+    fetch_separate_and_upload('s3://input/audio_example.mp3', 's3://output/audio_example.mp3', s3=mocked_cloud['s3'])
+    assert [e.key for e in mocked_cloud['output_bucket'].objects.all()] == ['audio_example.mp3']
+
+
+@pytest.mark.parametrize('url,expected_result', [
+    ('s3://bucket/key', ('bucket', 'key')),
+    ('s3://bucket/with/deep/nested/path.ext', ('bucket', 'with/deep/nested/path.ext')),
+])
+def test_parse_s3_url(url, expected_result):
+    assert parse_s3_url(url) == expected_result