diff --git a/app/services/llm_csv_parser.py b/app/services/llm_csv_parser.py
new file mode 100644
index 0000000..9d7a04c
--- /dev/null
+++ b/app/services/llm_csv_parser.py
@@ -0,0 +1,42 @@
+"""LLM-fallback CSV parser.
+
+When the deterministic Trading 212 parser (``csv_import.parse_t212_csv``)
+raises ``CSVImportError`` on an unrecognised format, this service kicks
+in:
+
+1. Detect the CSV dialect (delimiter, preamble offset).
+2. Compute a fingerprint of the normalised header row.
+3. Look up ``CsvFormatTemplate`` by fingerprint. On hit, replay the
+   cached column-mapping deterministically. On miss, ask the LLM for a
+   mapping, validate it, persist a new template, and apply it.
+
+The LLM sees only headers + the first 3-5 sample rows. It returns a
+column-mapping JSON, never transcribed numbers. The system never
+auto-promotes a learned format to a hand-written parser — the operator
+does that by inspecting collected ``sample_row`` values.
+"""
+from __future__ import annotations
+
+import hashlib
+
+from app.services.csv_import import CSVImportError
+
+
+class LLMParseError(CSVImportError):
+    """Raised when the LLM call fails or returns an unusable mapping.
+
+    Inherits from ``CSVImportError`` so route-level error handling can
+    treat both deterministic and LLM-path failures uniformly when
+    desired."""
+
+
+def _fingerprint(headers: list[str]) -> str:
+    """Stable hash of the header row.
+
+    Lowercases each header, strips surrounding whitespace, joins with
+    ``|`` (a character extremely unlikely to appear inside a real
+    header), and returns the sha256 hex digest. Whitespace/case drift
+    in the same broker's export does not change the fingerprint;
+    adding or removing a column does."""
+    normalised = "|".join(h.strip().lower() for h in headers)
+    return hashlib.sha256(normalised.encode("utf-8")).hexdigest()
diff --git a/tests/test_llm_csv_parser.py b/tests/test_llm_csv_parser.py
index 6ccf7ff..27223bb 100644
--- a/tests/test_llm_csv_parser.py
+++ b/tests/test_llm_csv_parser.py
@@ -29,3 +29,28 @@ def test_csv_format_template_model_columns():
     # Fingerprint is the cache key.
     assert cols["fingerprint"].unique is True
     assert cols["fingerprint"].nullable is False
+
+
+def test_fingerprint_stable_across_case_and_whitespace():
+    from app.services.llm_csv_parser import _fingerprint
+
+    a = _fingerprint(["Symbol", "Quantity", "Avg Price"])
+    b = _fingerprint(["symbol", "quantity", "avg price"])
+    c = _fingerprint(["  SYMBOL ", "Quantity", " AVG PRICE"])
+    assert a == b == c
+
+
+def test_fingerprint_differs_for_different_columns():
+    from app.services.llm_csv_parser import _fingerprint
+
+    a = _fingerprint(["Symbol", "Quantity"])
+    b = _fingerprint(["Symbol", "Quantity", "Avg Price"])
+    assert a != b
+
+
+def test_fingerprint_is_sha256_hex_64_chars():
+    from app.services.llm_csv_parser import _fingerprint
+
+    f = _fingerprint(["Symbol", "Quantity"])
+    assert len(f) == 64
+    assert all(c in "0123456789abcdef" for c in f)