diff --git a/src/diarize/__init__.py b/src/diarize/__init__.py
index fc1574f..68a907c 100644
--- a/src/diarize/__init__.py
+++ b/src/diarize/__init__.py
@@ -23,6 +23,7 @@
 from .clustering import cluster_speakers, estimate_speakers  # noqa: F401
 from .embeddings import extract_embeddings
 from .utils import (
+    DiarizeArtifacts,
     DiarizeResult,
     Segment,
     SpeakerEstimationDetails,  # noqa: F401
@@ -45,6 +46,7 @@ class _RawSegment(NamedTuple):
 __all__ = [
     "diarize",
     "DiarizeResult",
+    "DiarizeArtifacts",
     "Segment",
     "SpeakerEstimationDetails",
     "estimate_speakers",
@@ -129,6 +131,7 @@ def diarize(
     min_speakers: int = 1,
     max_speakers: int = 20,
     num_speakers: int | None = None,
+    return_artifacts: bool = False,
 ) -> DiarizeResult:
     """Run the full speaker diarization pipeline on an audio file.
 
@@ -146,6 +149,9 @@ def diarize(
         max_speakers: Maximum number of speakers for auto-detection.
         num_speakers: If set, skip auto-detection and use this exact
             number of speakers.
+        return_artifacts: If ``True``, populate
+            :attr:`DiarizeResult.artifacts` with speaker embeddings and
+            subsegments. Defaults to ``False``.
 
     Returns:
         :class:`DiarizeResult` containing segments, speaker info, and
@@ -202,6 +208,10 @@ def diarize(
         audio_path=audio_path_str,
         audio_duration=duration,
         estimation_details=estimation_details,
+        artifacts=DiarizeArtifacts(
+            embeddings=embeddings.tolist(),
+            subsegments=subsegments,
+        ) if return_artifacts else None,
     )
 
     logger.info(
diff --git a/src/diarize/utils.py b/src/diarize/utils.py
index 4a74188..7b5b40c 100644
--- a/src/diarize/utils.py
+++ b/src/diarize/utils.py
@@ -11,6 +11,7 @@
 from typing import Iterator
 
 import soundfile as sf
+
 from pydantic import BaseModel, Field, computed_field, model_validator
 
 logger = logging.getLogger(__name__)
@@ -21,6 +22,7 @@
     "SubSegment",
     "SpeakerEstimationDetails",
     "DiarizeResult",
+    "DiarizeArtifacts",
     "get_audio_duration",
     "format_timestamp",
 ]
@@ -121,6 +123,26 @@ class SpeakerEstimationDetails(BaseModel):
     cosine_sim_p10: float | None = None
 
 
+class DiarizeArtifacts(BaseModel):
+    """Internal pipeline artifacts, available when ``return_artifacts=True``.
+
+    All fields use plain Python types for clean serialization.
+
+    Attributes:
+        embeddings: Speaker embedding vectors (one per subsegment).
+        subsegments: Embedding windows used during extraction.
+    """
+
+    embeddings: list[list[float]] = Field(
+        default_factory=list,
+        description="Speaker embedding vectors as nested lists",
+    )
+    subsegments: list[SubSegment] = Field(
+        default_factory=list,
+        description="Embedding windows used during extraction",
+    )
+
+
 class DiarizeResult(BaseModel):
     """Result of speaker diarization.
 
@@ -145,6 +167,7 @@ class DiarizeResult(BaseModel):
     audio_path: str = ""
     audio_duration: float = Field(default=0.0, ge=0)
     estimation_details: SpeakerEstimationDetails | None = None
+    artifacts: DiarizeArtifacts | None = None
 
     @computed_field  # type: ignore[prop-decorator]
     @property