fix: relax cv experience parsing

Reject narrative verb-led bullets, allow short tech titles, and switch company prose checks to phrase-level matching so real employers like Made in Design are preserved.

src/job_research/profile/cv_extractor.py

@@ -35,6 +35,21 @@ EXPERIENCE_TITLE_STOPWORDS = {
     "because",
 }
 
+EXPERIENCE_TITLE_ACTION_STARTS = {
+    "built",
+    "created",
+    "delivered",
+    "designed",
+    "developed",
+    "deployed",
+    "implemented",
+    "improved",
+    "managed",
+    "migrated",
+    "maintained",
+    "worked",
+}
+
 EXPERIENCE_TITLE_LABELS = {
     "education",
     "experience",
@@ -79,6 +94,16 @@ EXPERIENCE_PROSE_MARKERS = {
     "worked",
 }
 
+EXPERIENCE_PROSE_COMPANY_PATTERNS = (
+    re.compile(r"\bau\s+sein\s+de\b", re.IGNORECASE),
+    re.compile(r"\b(before|after|during|while|since|because)\b", re.IGNORECASE),
+    re.compile(r"\b(joined|joining|moved|moving|relocated|relocating|worked|working)\b", re.IGNORECASE),
+    re.compile(
+        r"\bin\s+(?:paris|london|lyon|france|berlin|amsterdam|madrid|rome|marseille|bordeaux|toulouse|nantes|lille|grenoble|strasbourg|nice|rennes|montpellier|remote)\b",
+        re.IGNORECASE,
+    ),
+)
+
 EDUCATION_TITLE_KEYWORDS = {
     "bachelor",
     "degree",
@@ -293,9 +318,6 @@ def _looks_like_experience_title(title: str) -> bool:
         return False
 
     first_word = title_words[0]
-    if len(first_word) < 2:
-        return False
-
     normalized_title = " ".join(title_words).casefold()
     if normalized_title in EXPERIENCE_TITLE_LABELS:
         return False
@@ -303,16 +325,16 @@ def _looks_like_experience_title(title: str) -> bool:
     if first_word.casefold() in EXPERIENCE_TITLE_STOPWORDS:
         return False
 
+    if first_word.casefold() in EXPERIENCE_TITLE_ACTION_STARTS:
+        return False
+
+    if len(title_words) == 1 and len(first_word) < 2:
+        return False
+
     return not any(
         word.casefold() in EXPERIENCE_PROSE_MARKERS for word in title_words if word
     )
 
 
 def _looks_like_prose_company(company: str) -> bool:
-    company_words = [word.strip(".,;:!?()[]{}") for word in company.split()]
-    return any(
-        word
-        and word.lower() in EXPERIENCE_PROSE_MARKERS
-        and not (word.isupper() and len(word) <= 3)
-        for word in company_words
-    )
+    return any(pattern.search(company) for pattern in EXPERIENCE_PROSE_COMPANY_PATTERNS)

tests/profile/test_cv_extractor.py

@@ -169,7 +169,7 @@ def test_extract_cv_signals_parses_clear_titles_with_french_and_english_connecto
     ]
 
 
-def test_extract_cv_signals_accepts_lowercase_mixed_case_and_numeric_titles() -> None:
+def test_extract_cv_signals_accepts_lowercase_short_and_real_company_titles() -> None:
     text = dedent(
         """
         Tonio
@@ -177,6 +177,8 @@ def test_extract_cv_signals_accepts_lowercase_mixed_case_and_numeric_titles() ->
         data engineer at Company A
         iOS Engineer at Company A
         3D Artist at Studio
+        R Developer at Company A
+        Data Engineer at Made in Design
         """
     ).strip()
 
@@ -186,6 +188,8 @@ def test_extract_cv_signals_accepts_lowercase_mixed_case_and_numeric_titles() ->
         {"title": "data engineer", "company": "Company A"},
         {"title": "iOS Engineer", "company": "Company A"},
         {"title": "3D Artist", "company": "Studio"},
+        {"title": "R Developer", "company": "Company A"},
+        {"title": "Data Engineer", "company": "Made in Design"},
     ]
 
 
@@ -219,6 +223,21 @@ def test_extract_cv_signals_rejects_label_like_lines_without_colons() -> None:
     assert extracted["experience_entries"] == []
 
 
+def test_extract_cv_signals_rejects_narrative_bullet_experience_lines() -> None:
+    text = dedent(
+        """
+        Tonio
+        Location: France
+        Implemented data pipelines at Airbnb
+        Designed dashboards at Company A
+        """
+    ).strip()
+
+    extracted = extract_cv_signals(text)
+
+    assert extracted["experience_entries"] == []
+
+
 def test_extract_cv_signals_ignores_in_paris_prose_tail() -> None:
     text = dedent(
         """