diff --git a/retry_mechanism_analysis/src/retry_mechanism_analysis/core.py b/retry_mechanism_analysis/src/retry_mechanism_analysis/core.py
new file mode 100644
index 0000000..5217a8e
--- /dev/null
+++ b/retry_mechanism_analysis/src/retry_mechanism_analysis/core.py
@@ -0,0 +1,60 @@
+import pandas as pd
+import logging
+from typing import List, Dict, Any, TypedDict
+
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+class RetryAnalysis(TypedDict):
+    parallelism_level: int
+    p50: float
+    p95: float
+    p99: float
+
+
+def _validate_log_data(log_data: List[Dict[str, Any]]) -> None:
+    if not isinstance(log_data, list):
+        raise ValueError("log_data must be a list of dictionaries.")
+    for entry in log_data:
+        if not isinstance(entry, dict):
+            raise ValueError("Each log entry must be a dictionary.")
+        required_keys = {"parallelism_level", "retry_latency_ms"}
+        missing = required_keys - set(entry.keys())
+        if missing:
+            raise ValueError(f"Missing keys in log entry: {missing}")
+        if not isinstance(entry["parallelism_level"], int):
+            raise ValueError("'parallelism_level' must be an integer.")
+        if not isinstance(entry["retry_latency_ms"], (int, float)):
+            raise ValueError("'retry_latency_ms' must be numeric.")
+
+
+def compare_retry_overhead(log_data: List[Dict[str, Any]]) -> List[RetryAnalysis]:
+    """Vergleicht Retry-Overheads über verschiedene Parallelitätslevels und liefert aggregierte Metriken."""
+    _validate_log_data(log_data)
+
+    logger.info("Start comparing retry overheads across parallelism levels.")
+
+    df = pd.DataFrame(log_data)
+    if df.empty:
+        raise ValueError("log_data is empty; no data to analyze.")
+
+    grouped = df.groupby("parallelism_level")["retry_latency_ms"]
+
+    results: List[RetryAnalysis] = []
+    for level, latencies in grouped:
+        p50 = float(latencies.quantile(0.5))
+        p95 = float(latencies.quantile(0.95))
+        p99 = float(latencies.quantile(0.99))
+        analysis: RetryAnalysis = {
+            "parallelism_level": int(level),
+            "p50": p50,
+            "p95": p95,
+            "p99": p99,
+        }
+        results.append(analysis)
+        logger.debug(f"Computed metrics for parallelism {level}: {analysis}")
+
+    logger.info("Completed retry overhead comparison.")
+    return results
\ No newline at end of file