diff --git a/core/inference_client.py b/core/inference_client.py
index 318c9a7..6ecb09c 100644
--- a/core/inference_client.py
+++ b/core/inference_client.py
@@ -24,6 +24,8 @@ class InferenceClient:
     def __init__(self):
         self.server_process: Optional[subprocess.Popen] = None
         self._server_lock = threading.Lock()
+        self.log_file = None
+        self.log_file_path = None
     
     def start_server(self):
         """Start the inference server process."""
@@ -52,28 +54,54 @@ class InferenceClient:
                             server_env[key] = value
                 print(f"[FaceMask] Loaded environment from: {env_file}")
 
-            # Ensure PYTHONPATH includes project root
-            pythonpath = server_env.get('PYTHONPATH', '')
-            if pythonpath:
-                server_env['PYTHONPATH'] = f"{root_dir}:{pythonpath}"
-            else:
-                server_env['PYTHONPATH'] = root_dir
+            # Clean PYTHONPATH to avoid conflicts with Nix Python packages
+            # Only include project root to allow local imports
+            server_env['PYTHONPATH'] = root_dir
+
+            # Remove Python-related environment variables that might cause conflicts
+            # These can cause venv to import packages from Nix instead of venv
+            env_vars_to_remove = [
+                'PYTHONUNBUFFERED',
+                '__PYVENV_LAUNCHER__',  # macOS venv variable
+                'VIRTUAL_ENV',  # Will be set by venv's Python automatically
+            ]
+            for var in env_vars_to_remove:
+                server_env.pop(var, None)
 
             # If there's a venv in the project, add it to PATH
             venv_bin = os.path.join(root_dir, ".venv", "bin")
             if os.path.isdir(venv_bin):
+                # Build a clean PATH with venv first, then essential system paths
+                # Filter out any Nix Python-specific paths to avoid version conflicts
                 current_path = server_env.get('PATH', '')
-                server_env['PATH'] = f"{venv_bin}:{current_path}"
+                path_entries = current_path.split(':')
+
+                # Filter out Nix Python 3.11 paths
+                filtered_paths = [
+                    p for p in path_entries
+                    if not ('/python3.11/' in p.lower() or '/python3-3.11' in p.lower())
+                ]
+
+                # Reconstruct PATH with venv first
+                clean_path = ':'.join([venv_bin] + filtered_paths)
+                server_env['PATH'] = clean_path
                 print(f"[FaceMask] Using venv from: {venv_bin}")
 
+            # Prepare log file for server output
+            import tempfile
+            log_dir = tempfile.gettempdir()
+            self.log_file_path = os.path.join(log_dir, "facemask_server.log")
+            self.log_file = open(self.log_file_path, 'w', buffering=1)  # Line buffered
+            print(f"[FaceMask] Server log: {self.log_file_path}")
+
             # Start process with 'python' command (will use venv if PATH is set correctly)
             self.server_process = subprocess.Popen(
-                ["python", server_script],
+                ["python", "-u", server_script],  # -u for unbuffered output
                 cwd=root_dir,
                 text=True,
                 env=server_env,
-                stdout=subprocess.PIPE,  # Capture stdout
-                stderr=subprocess.PIPE,  # Capture stderr
+                stdout=self.log_file,  # Write to log file
+                stderr=subprocess.STDOUT,  # Merge stderr into stdout
                 preexec_fn=os.setsid,  # Create new process group
             )
 
@@ -85,22 +113,23 @@ class InferenceClient:
 
                 # Check if process died
                 if self.server_process.poll() is not None:
-                    # Capture and display error output
-                    try:
-                        stdout, stderr = self.server_process.communicate(timeout=1)
-                    except subprocess.TimeoutExpired:
-                        stdout, stderr = "", ""
-
+                    # Read error output from log file
                     error_msg = f"Server failed to start (exit code: {self.server_process.returncode})"
                     print(f"[FaceMask] ERROR: {error_msg}")
 
-                    if stdout and stdout.strip():
-                        print("[FaceMask] Server stdout:")
-                        print(stdout.strip())
-
-                    if stderr and stderr.strip():
-                        print("[FaceMask] Server stderr:")
-                        print(stderr.strip())
+                    try:
+                        if self.log_file:
+                            self.log_file.close()
+                        with open(self.log_file_path, 'r') as f:
+                            log_content = f.read()
+                            if log_content.strip():
+                                print("[FaceMask] Server log:")
+                                # Show last 50 lines
+                                lines = log_content.strip().split('\n')
+                                for line in lines[-50:]:
+                                    print(line)
+                    except Exception as e:
+                        print(f"[FaceMask] Could not read log file: {e}")
 
                     self.server_process = None
                     raise RuntimeError(error_msg)
@@ -108,19 +137,21 @@ class InferenceClient:
                 time.sleep(0.5)
 
             # If we get here, startup timed out
-            # Try to capture any partial output
-            if self.server_process:
-                try:
-                    # Non-blocking read with short timeout
-                    stdout, stderr = self.server_process.communicate(timeout=0.1)
-                    if stdout and stdout.strip():
-                        print("[FaceMask] Server stdout (partial):")
-                        print(stdout.strip())
-                    if stderr and stderr.strip():
-                        print("[FaceMask] Server stderr (partial):")
-                        print(stderr.strip())
-                except subprocess.TimeoutExpired:
-                    pass  # Process still running but not responding
+            print("[FaceMask] Server startup timed out")
+
+            # Try to read partial log
+            try:
+                if self.log_file:
+                    self.log_file.close()
+                with open(self.log_file_path, 'r') as f:
+                    log_content = f.read()
+                    if log_content.strip():
+                        print("[FaceMask] Server log (partial):")
+                        lines = log_content.strip().split('\n')
+                        for line in lines[-30:]:
+                            print(line)
+            except Exception:
+                pass
 
             raise RuntimeError("Server startup timed out")
     
@@ -136,6 +167,14 @@ class InferenceClient:
                     pass
                 finally:
                     self.server_process = None
+
+            # Close log file
+            if self.log_file:
+                try:
+                    self.log_file.close()
+                except Exception:
+                    pass
+                self.log_file = None
     
     def is_server_running(self) -> bool:
         """Check if server is responding."""