diff --git a/tests/test_llm.py b/tests/test_llm.py
new file mode 100644
index 0000000..579c7e5
--- /dev/null
+++ b/tests/test_llm.py
@@ -0,0 +1,190 @@
+"""
+Unit tests for the LLM class (src/llm.py).
+
+Tests cover:
+- build_prompt()         → prompt contains the right field and transcript
+- handle_plural_values() → splits "val1; val2" into ["val1", "val2"]
+- add_response_to_json() → stores None for "-1", stores value otherwise
+- main_loop()            → mocked Ollama call, builds JSON correctly
+"""
+
+import pytest
+from unittest.mock import patch, MagicMock
+from src.llm import LLM
+
+
+# ─────────────────────────────────────────────────────────────
+# build_prompt() tests
+# ─────────────────────────────────────────────────────────────
+
+def test_build_prompt_contains_field_name():
+    """Prompt must mention the target field so the LLM knows what to extract."""
+    llm = LLM(transcript_text="John Doe was the employee.", target_fields={"name": ""})
+    prompt = llm.build_prompt("name")
+    assert "name" in prompt
+
+
+def test_build_prompt_contains_transcript_text():
+    """Prompt must include the full transcript so the LLM has context."""
+    transcript = "Station 5 responded to a fire at 123 Main Street."
+    llm = LLM(transcript_text=transcript, target_fields={"location": ""})
+    prompt = llm.build_prompt("location")
+    assert transcript in prompt
+
+
+def test_build_prompt_returns_string():
+    """build_prompt should always return a string."""
+    llm = LLM(transcript_text="Some text.", target_fields={"date": ""})
+    result = llm.build_prompt("date")
+    assert isinstance(result, str)
+
+
+def test_build_prompt_contains_both_field_and_transcript():
+    """Full check: both field name and transcript must appear in the prompt."""
+    llm = LLM(transcript_text="John Doe", target_fields={"name": ""})
+    prompt = llm.build_prompt("name")
+    assert "name" in prompt
+    assert "John Doe" in prompt
+
+
+# ─────────────────────────────────────────────────────────────
+# handle_plural_values() tests
+# ─────────────────────────────────────────────────────────────
+
+def test_handle_plural_values_splits_two_values():
+    """Standard case: two values separated by semicolon."""
+    llm = LLM()
+    result = llm.handle_plural_values("Engine 12; Engine 14")
+    assert result == ["Engine 12", "Engine 14"]
+
+
+def test_handle_plural_values_strips_leading_whitespace():
+    """Values after semicolons often have leading spaces — must be stripped."""
+    llm = LLM()
+    result = llm.handle_plural_values("Alice;  Bob;  Charlie")
+    # The method only strips the second element onwards (based on current impl)
+    assert "Alice" in result
+    assert any("Bob" in v for v in result)
+
+
+def test_handle_plural_values_raises_on_no_semicolon():
+    """If no semicolon is present, raise ValueError."""
+    llm = LLM()
+    with pytest.raises(ValueError):
+        llm.handle_plural_values("Engine 12")
+
+
+def test_handle_plural_values_returns_list():
+    """Return type must always be a list."""
+    llm = LLM()
+    result = llm.handle_plural_values("A; B")
+    assert isinstance(result, list)
+
+
+# ─────────────────────────────────────────────────────────────
+# add_response_to_json() tests
+# ─────────────────────────────────────────────────────────────
+
+def test_add_response_ignores_minus_one():
+    """When LLM returns '-1' (field not found), JSON value should be None."""
+    llm = LLM(target_fields={}, json={})
+    llm.add_response_to_json("name", "-1")
+    assert llm._json["name"] is None
+
+
+def test_add_response_stores_valid_value():
+    """When LLM returns a real value, it should be stored in the JSON dict."""
+    llm = LLM(target_fields={}, json={})
+    llm.add_response_to_json("name", "John Doe")
+    assert llm._json["name"] == "John Doe"
+
+
+def test_add_response_strips_quotes_from_value():
+    """LLM responses sometimes have extra quotes — they should be removed."""
+    llm = LLM(target_fields={}, json={})
+    llm.add_response_to_json("name", '"John Doe"')
+    assert llm._json["name"] == "John Doe"
+
+
+def test_add_response_handles_plural_values_with_semicolon():
+    """When value contains ';', it should be split into a list."""
+    llm = LLM(target_fields={}, json={})
+    llm.add_response_to_json("engines", "Engine 12; Engine 14")
+    assert isinstance(llm._json["engines"], list)
+
+
+def test_add_response_creates_new_field_if_not_exists():
+    """If the field doesn't exist in JSON yet, it should be created."""
+    llm = LLM(target_fields={}, json={})
+    llm.add_response_to_json("department", "Cal Fire")
+    assert "department" in llm._json
+
+
+# ─────────────────────────────────────────────────────────────
+# main_loop() tests — mocked Ollama API
+# ─────────────────────────────────────────────────────────────
+
+def test_main_loop_calls_ollama_once_per_field():
+    """main_loop() should call the Ollama API once per field in target_fields."""
+    llm = LLM(
+        transcript_text="John Doe works in IT.",
+        target_fields={"name": "", "department": ""},
+        json={"name": "", "department": ""},
+    )
+
+    mock_response = MagicMock()
+    mock_response.json.return_value = {"response": "extracted_value"}
+    mock_response.raise_for_status = MagicMock()
+
+    with patch("requests.post", return_value=mock_response) as mock_post:
+        llm.main_loop()
+        # Should be called once for each field (2 fields = 2 calls)
+        assert mock_post.call_count == 2
+
+
+def test_main_loop_returns_self():
+    """main_loop() should return the LLM instance (for chaining)."""
+    llm = LLM(
+        transcript_text="Some incident text.",
+        target_fields={"name": ""},
+        json={"name": ""},
+    )
+
+    mock_response = MagicMock()
+    mock_response.json.return_value = {"response": "John"}
+    mock_response.raise_for_status = MagicMock()
+
+    with patch("requests.post", return_value=mock_response):
+        result = llm.main_loop()
+        assert result is llm
+
+
+def test_main_loop_builds_correct_json_from_ollama_response():
+    """main_loop() should populate _json with the Ollama response values."""
+    llm = LLM(
+        transcript_text="The employee is John Doe.",
+        target_fields={"name": ""},
+        json={"name": ""},
+    )
+
+    mock_response = MagicMock()
+    mock_response.json.return_value = {"response": "John Doe"}
+    mock_response.raise_for_status = MagicMock()
+
+    with patch("requests.post", return_value=mock_response):
+        llm.main_loop()
+        assert llm._json.get("name") == "John Doe"
+
+
+def test_main_loop_raises_on_ollama_connection_error():
+    """If Ollama is not running, main_loop() should raise a ConnectionError."""
+    import requests as req
+    llm = LLM(
+        transcript_text="Some text.",
+        target_fields={"name": ""},
+        json={"name": ""},
+    )
+
+    with patch("requests.post", side_effect=req.exceptions.ConnectionError()):
+        with pytest.raises(ConnectionError):
+            llm.main_loop()