diff --git a/src/smithy/rcv.py b/src/smithy/rcv.py
index 4c99c74..c9f556e 100644
--- a/src/smithy/rcv.py
+++ b/src/smithy/rcv.py
@@ -3,7 +3,7 @@ import rustworkx as rwx
 from itertools import combinations
 
 
-def pmg_from_rcv(ballots: pl.DataFrame) -> rwx.PyDiGraph:
+def _pmg_from_rcv(ballots: pl.DataFrame) -> rwx.PyDiGraph:
     """
     Build a pairwise majority winner graph from a box of Ranked-Choice Ballots.
 
@@ -51,3 +51,48 @@ def pmg_from_rcv(ballots: pl.DataFrame) -> rwx.PyDiGraph:
             pmg.add_edge(nodes[b], nodes[a], b_wins - a_wins)
 
     return pmg
+
+def pmg_from_rcv(ballots: pl.DataFrame) -> rwx.PyDiGraph:
+    """
+    Build a pairwise majority winner graph from a box of Ranked-Choice Ballots.
+
+    parameters
+    ---
+    rcv_ballots : pl.DataFrame
+        A Polars DataFrame representing ballots. Each column is a candidate and each
+        row is is a voter's ranking of the candidates. Lower numbers indicate higher
+        preference (1 = top-choice).
+
+    returns
+    ---
+    nodes: dict[str, int]
+        A dictionary of candidate names to associated node ids.
+
+    pwm_graph: rwx.PyDiGraph
+        A pairwise majority winner graph whose nodes correspond to candidates and
+        (directed) edges show which candidates they beat pairwise.
+    """
+    candidates = ballots.columns
+
+    pmg = rwx.PyDiGraph()
+    nodes = {c: pmg.add_node(c) for c in candidates}
+
+    compressed = ballots.group_by(ballots.columns).len().rename({"len": "count"})
+    counts = compressed["count"].to_numpy()
+    
+    arr = compressed.drop("count").to_numpy()
+    results = ((arr[:, :, None] < arr[:, None, :]) * counts[:, None, None]).sum(axis=0)
+
+    for i, a in enumerate(candidates):
+        for j in range(i + 1, len(candidates)):
+            b = candidates[j]
+            
+            a_wins = results[i, j]
+            b_wins = results[j, i]
+            
+            if a_wins > b_wins:
+                pmg.add_edge(nodes[a], nodes[b], int(a_wins - b_wins))
+            elif b_wins > a_wins:
+                pmg.add_edge(nodes[b], nodes[a], int(b_wins - a_wins))
+
+    return pmg
\ No newline at end of file