link added

app4.py

@@ -1,396 +1,396 @@
-# streamlit_app.py
-import streamlit as st
-import pandas as pd
-import numpy as np
-import joblib
-from typing import Tuple, Dict, Any, List
-
-st.set_page_config(page_title="USA Salary — Synthetic/Hybrid Prediction", layout="wide")
-
-# --- Top nav to teammates' apps ---
-c1, c2, c3 = st.columns(3)
-with c1:
-    st.link_button("Hamna", "https://example.com/hamna")   # TODO: replace with real link
-with c2:
-    st.link_button("Mahesh", "https://example.com/mahesh") # TODO: replace with real link
-with c3:
-    st.link_button("Tian", "https://example.com/tian")     # TODO: replace with real link
-
-st.title("USA Salary — Predict with Hybrid/Synthetic Inputs")
-
-# =================== Load assets ===================
-@st.cache_resource(show_spinner=False)
-def _load_assets():
-    pipe = joblib.load("final_xgbr_usa_model.pkl")
-    usa_2024 = pd.read_csv("usa_salary_data.csv")
-    usa_2025 = pd.read_csv("2025_survey.csv")
-    return pipe, usa_2024, usa_2025
-
-pipe, usa_data, usa_25 = _load_assets()
-LABEL = "CompTotal"
-
-# =================== RNG utilities ===================
-def _ensure_rng():
-    if "rng_seedseq" not in st.session_state:
-        st.session_state.rng_seedseq = np.random.SeedSequence()
-    child = st.session_state.rng_seedseq.spawn(1)[0]
-    return np.random.default_rng(child)
-
-def new_rng():
-    st.session_state.rng_seedseq = st.session_state.rng_seedseq.spawn(1)[0]
-    return np.random.default_rng(st.session_state.rng_seedseq)
-
-# =================== Precompute dropdown choices (ONCE) ===================
-@st.cache_resource(show_spinner=False)
-def _precompute_choices(usa_2024: pd.DataFrame, usa_2025: pd.DataFrame, label: str) -> Dict[str, List]:
-    """
-    Build a dict of column -> list of allowed choices.
-    - Categorical: top-k frequent from union(2024, 2025)
-    - Numeric: percentile grid + most frequent rounded values
-    """
-    CHOICES: Dict[str, List] = {}
-    both = pd.concat([usa_2024, usa_2025.reindex(columns=usa_2024.columns, fill_value=np.nan)], axis=0, ignore_index=True)
-    for col in usa_2024.columns:
-        if col == label:  # never expose label to edit
-            continue
-
-        s = both[col]
-        if pd.api.types.is_numeric_dtype(usa_2024[col]):
-            s_num = pd.to_numeric(s, errors="coerce").dropna()
-            if len(s_num) == 0:
-                CHOICES[col] = [0]
-                continue
-
-            # Percentile grid (5th..95th) + median
-            q_list = [5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 95]
-            qs = np.percentile(s_num, q_list).tolist()
-
-            # Common rounded values (to nearest 1000) by frequency
-            rounded = (np.round(s_num / 1000) * 1000).astype(int)
-            top_round = rounded.value_counts().head(10).index.astype(int).tolist()
-
-            # Merge, deduplicate, sort, and keep as Python numbers (not numpy types)
-            merged = sorted(set(int(v) for v in qs + top_round))
-            CHOICES[col] = merged[:50]  # cap length
-        else:
-            s_cat = s.astype(str).replace({"nan": None})
-            s_cat = s_cat.dropna()
-            if len(s_cat) == 0:
-                CHOICES[col] = [""]
-                continue
-            topk = s_cat.value_counts().head(40).index.tolist()
-            CHOICES[col] = topk
-    return CHOICES
-
-CHOICES_DICT = _precompute_choices(usa_data, usa_25, LABEL)
-
-# =================== Core sampling funcs (your logic) ===================
-def _sample_from_2024(colname: str, df_2024: pd.DataFrame, rng: np.random.Generator):
-    series = df_2024[colname].dropna()
-    if series.empty:
-        if pd.api.types.is_numeric_dtype(df_2024[colname]):
-            return 0
-        return ""
-    return series.sample(1, random_state=rng.integers(0, 10_000)).iloc[0]
-
-def build_synthetic_row_with_trace(
-    usa_25: pd.DataFrame,
-    usa_2024: pd.DataFrame,
-    label: str = "CompTotal",
-    rng: np.random.Generator | None = None
-) -> Tuple[pd.DataFrame, float | None, Dict[str, str], Dict[str, Any]]:
-    if rng is None:
-        rng = _ensure_rng()
-
-    expected_features = [c for c in usa_2024.columns if c != label]
-    row25 = usa_25.sample(1, random_state=rng.integers(0, 10_000)).iloc[0]
-
-    synthetic = {}
-    source_info: Dict[str, str] = {}
-
-    for col in expected_features:
-        use_25_val = False
-        val = None
-
-        if col in row25.index:
-            val = row25[col]
-            if pd.api.types.is_numeric_dtype(usa_2024[col]):
-                val = pd.to_numeric(val, errors="coerce")
-            if not pd.isna(val):
-                use_25_val = True
-
-        if not use_25_val:
-            val = _sample_from_2024(col, usa_2024, rng)
-            source_info[col] = "2024"
-        else:
-            source_info[col] = "2025"
-
-        if pd.api.types.is_numeric_dtype(usa_2024[col]):
-            val = pd.to_numeric(val, errors="coerce")
-            if pd.isna(val):
-                val = _sample_from_2024(col, usa_2024, rng)
-                source_info[col] = "2024"
-        else:
-            if pd.isna(val):
-                val = _sample_from_2024(col, usa_2024, rng)
-                source_info[col] = "2024"
-            if not isinstance(val, str):
-                val = str(val)
-
-        synthetic[col] = val
-
-    X_one = pd.DataFrame([synthetic], columns=expected_features)
-
-    y_true = None
-    if label in row25.index:
-        y_true = pd.to_numeric(row25[label], errors="coerce")
-        if pd.isna(y_true):
-            y_true = None
-
-    used_from_2025 = sum(v == "2025" for v in source_info.values())
-    used_from_2024 = sum(v == "2024" for v in source_info.values())
-    total = len(source_info)
-    report = dict(
-        total_features=total,
-        n_2025=used_from_2025,
-        n_2024=used_from_2024,
-        pct_2025=0.0 if total == 0 else used_from_2025 / total * 100,
-        pct_2024=0.0 if total == 0 else used_from_2024 / total * 100,
-        filled_cols=[c for c, s in source_info.items() if s == "2024"]
-    )
-    return X_one, y_true, source_info, report
-
-def random_sample_row_2024(
-    df_2024: pd.DataFrame,
-    label: str,
-    rng: np.random.Generator
-) -> pd.DataFrame:
-    expected_features = [c for c in df_2024.columns if c != label]
-    sampled = {}
-    for col in expected_features:
-        sampled[col] = _sample_from_2024(col, df_2024, rng)
-    return pd.DataFrame([sampled], columns=expected_features)
-
-# =================== Session state ===================
-if "history" not in st.session_state:
-    st.session_state.history = []  # list of dicts: pred, truth, abs_err, pct_err
-
-if "prepared" not in st.session_state:
-    st.session_state.prepared = None  # dict: X, y_true, info, report, mode
-
-# =================== Sidebar ===================
-st.sidebar.header("Controls")
-mode = st.sidebar.radio(
-    "Choose input mode:",
-    ["Hybrid (Random 2025 + fill from 2024)", "Pure 2024 synthetic"],
-    index=0
-)
-
-if st.sidebar.button("Reload new random data"):
-    rng = new_rng()
-    if mode.startswith("Hybrid"):
-        X_one, y_true, src_info, rep = build_synthetic_row_with_trace(usa_25, usa_data, label=LABEL, rng=rng)
-        st.session_state.prepared = dict(X=X_one, y_true=y_true, info=src_info, report=rep, mode="hybrid")
-    else:
-        X_2024 = random_sample_row_2024(usa_data, LABEL, rng)
-        st.session_state.prepared = dict(X=X_2024, y_true=None, info={}, report={}, mode="pure2024")
-    st.toast("New random input prepared.", icon="✅")
-
-# =================== Explanation ===================
-with st.expander("What is happening here? "):
-    st.markdown(
-        """
-**Goal:** Create one model-ready row and predict **CompTotal(Annual Income)**.
-**dataset** trained on: USA data from 2024 stackoverflow survey.
-            predict on: synthetic/hybrid data mixing 2025 & 2024.
-**Modes:**
-1) **Hybrid** — Pick a random 2025 respondent; for any missing required feature, fill using a value sampled from the 2024 distribution for that column.
-2) **Pure 2024** — Build an entirely synthetic row, sampling every feature from the 2024 distribution.
-
-**Editing:**  
-Below, you can adjust the *current* row via dropdowns. All dropdown choices were **precomputed once at startup** from the union of 2024 & 2025 data (to keep the app fast).  
-- **Categorical**: most frequent categories (+ always includes the current value).  
-- **Numeric**: percentile grid & common rounded values (+ current value).
-
-Click **Submit & Predict** to see predicted vs. true (if available), absolute and percentage errors, and running averages.
-        """
-    )
-
-st.markdown("---")
-
-# =================== Prepare first candidate if needed ===================
-def _prepare_candidate_if_needed():
-    if st.session_state.prepared is None:
-        rng = _ensure_rng()
-        if mode.startswith("Hybrid"):
-            X_one, y_true, src_info, rep = build_synthetic_row_with_trace(usa_25, usa_data, label=LABEL, rng=rng)
-            st.session_state.prepared = dict(X=X_one, y_true=y_true, info=src_info, report=rep, mode="hybrid")
-        else:
-            X_2024 = random_sample_row_2024(usa_data, LABEL, rng)
-            st.session_state.prepared = dict(X=X_2024, y_true=None, info={}, report={}, mode="pure2024")
-
-_prepare_candidate_if_needed()
-
-# =================== Helper: coerce edited values to correct dtype ===================
-def _coerce_value(col: str, val):
-    if pd.api.types.is_numeric_dtype(usa_data[col]):
-        # Accept numbers that might come from selectbox as str/float
-        try:
-            return pd.to_numeric(val)
-        except Exception:
-            return np.nan
-    else:
-        return "" if val is None else str(val)
-
-
-# =================== Predict & History ===================
-left, right = st.columns([1.1, 0.9], gap="large")
-
-with left:
-    st.subheader("Submit a prediction")
-
-    # Preview current candidate
-    if st.session_state.prepared is not None:
-        curr = st.session_state.prepared
-        mode_tag = "Hybrid (2025+2024)" if curr["mode"] == "hybrid" else "Pure 2024"
-        st.caption(f"Next input mode: **{mode_tag}**")
-
-        with st.expander("Show current input row (after your edits)"):
-            st.dataframe(curr["X"].T.rename(columns={0: "value"}))
-
-        if curr["mode"] == "hybrid":
-            rep = curr["report"]
-            st.caption(
-                f"Data completion: **{rep['n_2025']}** from 2025 "
-                f"({rep['pct_2025']:.1f}%); **{rep['n_2024']}** from 2024 "
-                f"({rep['pct_2024']:.1f}%)"
-            )
-
-    submitted = st.button("Submit & Predict", type="primary", use_container_width=True)
-
-    if submitted and st.session_state.prepared is not None:
-        curr = st.session_state.prepared
-        X_one = curr["X"]
-        y_true = curr["y_true"]
-
-        # (Optional) Final dtype alignment just before predict
-        try:
-            # Align numeric dtypes to training schema to be safe
-            for col in X_one.columns:
-                if pd.api.types.is_numeric_dtype(usa_data[col]):
-                    X_one[col] = pd.to_numeric(X_one[col], errors="coerce")
-                else:
-                    X_one[col] = X_one[col].astype(str).fillna("")
-        except Exception:
-            pass
-
-        try:
-            y_pred = float(pipe.predict(X_one)[0])
-        except Exception as e:
-            st.error(f"Prediction failed: {e}")
-            y_pred = None
-
-        if y_pred is not None:
-            st.success(f"**Predicted CompTotal:** {y_pred:,.0f} USD")
-
-            if y_true is not None:
-                st.info(f"**2025 true:** {y_true:,.0f} USD")
-
-                abs_err = abs(y_pred - y_true)
-                pct_err = abs_err / y_true * 100 if y_true != 0 else np.nan
-
-                st.write(f"**Absolute error:** {abs_err:,.0f} USD")
-                st.write(f"**Percentage error:** {pct_err:.2f}%")
-
-                st.session_state.history.append(
-                    dict(pred=y_pred, truth=y_true, abs_err=abs_err, pct_err=pct_err)
-                )
-            else:
-                st.warning("No ground-truth value available for this input (pure 2024 synthetic).")
-
-        # Prepare a new random candidate (and your edits will apply to the new one next round)
-        rng = new_rng()
-        if mode.startswith("Hybrid"):
-            X_one2, y_true2, src_info2, rep2 = build_synthetic_row_with_trace(usa_25, usa_data, label=LABEL, rng=rng)
-            st.session_state.prepared = dict(X=X_one2, y_true=y_true2, info=src_info2, report=rep2, mode="hybrid")
-        else:
-            X_2024b = random_sample_row_2024(usa_data, LABEL, rng)
-            st.session_state.prepared = dict(X=X_2024b, y_true=None, info={}, report={}, mode="pure2024")
-        st.toast("New random input prepared.", icon="✨")
-
-with right:
-    st.subheader("Results history")
-    if len(st.session_state.history) == 0:
-        st.write("No submissions yet.")
-    else:
-        hist_df = pd.DataFrame(st.session_state.history)
-        st.dataframe(
-            hist_df.style.format({"pred": "{:,.0f}", "truth": "{:,.0f}", "abs_err": "{:,.0f}", "pct_err": "{:.2f}"}),
-            use_container_width=True
-        )
-
-        valid = hist_df.dropna(subset=["truth"])
-        if len(valid) > 0:
-            mae = valid["abs_err"].mean()
-            mape = valid["pct_err"].mean()
-            st.metric(label="Mean Absolute Error (USD)", value=f"{mae:,.0f}")
-            st.metric(label="Mean Absolute Percentage Error", value=f"{mape:.2f}%")
-        else:
-            st.write("No entries with ground truth yet.")
-
-st.markdown("---")
-st.caption("Tip: Use the sidebar **Reload new random data** to resample without submitting.")
-# =================== Editable input row UI ===================
-st.subheader("Edit current input (optional)")
-if st.session_state.prepared is not None:
-    curr = st.session_state.prepared
-    X_row = curr["X"].iloc[0].copy()
-
-    # Two tabs for readability
-    tab_cat, tab_num = st.tabs(["Categorical features", "Numeric features"])
-
-    # Build lists
-    cat_cols = [c for c in X_row.index if not pd.api.types.is_numeric_dtype(usa_data[c])]
-    num_cols = [c for c in X_row.index if pd.api.types.is_numeric_dtype(usa_data[c])]
-
-    with tab_cat:
-        st.caption("Pick from common categories (precomputed). Your current value is preselected.")
-        for col in cat_cols:
-            choices = CHOICES_DICT.get(col, [])
-            # ensure current value is present
-            curr_val = "" if pd.isna(X_row[col]) else str(X_row[col])
-            if curr_val not in choices and curr_val != "":
-                choices = [curr_val] + choices
-            sel = st.selectbox(
-                label=col,
-                options=choices if len(choices) > 0 else [""],
-                index=0 if len(choices) == 0 else (choices.index(curr_val) if curr_val in choices else 0),
-                key=f"edit_cat_{col}",
-            )
-            X_row[col] = _coerce_value(col, sel)
-
-    with tab_num:
-        st.caption("Pick typical numeric values (percentiles/rounded) or keep current.")
-        for col in num_cols:
-            choices = CHOICES_DICT.get(col, [])
-            curr_val = X_row[col]
-            # ensure current value is present and cast to int for display if close to int
-            if pd.isna(curr_val):
-                curr_val = choices[0] if len(choices) else 0
-            # make sure current is in choices
-            if len(choices) == 0:
-                choices = [curr_val]
-            elif curr_val not in choices:
-                choices = [curr_val] + choices
-            sel = st.selectbox(
-                label=col,
-                options=choices,
-                index=choices.index(curr_val) if curr_val in choices else 0,
-                key=f"edit_num_{col}",
-            )
-            X_row[col] = _coerce_value(col, sel)
-
-    # Save edits back
-    st.session_state.prepared["X"].iloc[0] = X_row
-
-
+# streamlit_app.py
+import streamlit as st
+import pandas as pd
+import numpy as np
+import joblib
+from typing import Tuple, Dict, Any, List
+
+st.set_page_config(page_title="USA Salary — Synthetic/Hybrid Prediction", layout="wide")
+
+# --- Top nav to teammates' apps ---
+c1, c2, c3 = st.columns(3)
+with c1:
+    st.link_button("Hamna", "https://project-ytgzknejj7pcncnqxbcwua.streamlit.app/")   # TODO: replace with real link
+with c2:
+    st.link_button("Mahesh", "https://example.com/mahesh") # TODO: replace with real link
+with c3:
+    st.link_button("Tian", "https://www.canva.com/design/DAG1MBF-wU0/8rWZE_GCqcqBnNCoEECpGw/edit?utm_content=DAG1MBF-wU0&utm_campaign=designshare&utm_medium=link2&utm_source=sharebutton")     # TODO: replace with real link
+
+st.title("USA Salary — Predict with Hybrid/Synthetic Inputs")
+
+# =================== Load assets ===================
+@st.cache_resource(show_spinner=False)
+def _load_assets():
+    pipe = joblib.load("final_xgbr_usa_model.pkl")
+    usa_2024 = pd.read_csv("usa_salary_data.csv")
+    usa_2025 = pd.read_csv("2025_survey.csv")
+    return pipe, usa_2024, usa_2025
+
+pipe, usa_data, usa_25 = _load_assets()
+LABEL = "CompTotal"
+
+# =================== RNG utilities ===================
+def _ensure_rng():
+    if "rng_seedseq" not in st.session_state:
+        st.session_state.rng_seedseq = np.random.SeedSequence()
+    child = st.session_state.rng_seedseq.spawn(1)[0]
+    return np.random.default_rng(child)
+
+def new_rng():
+    st.session_state.rng_seedseq = st.session_state.rng_seedseq.spawn(1)[0]
+    return np.random.default_rng(st.session_state.rng_seedseq)
+
+# =================== Precompute dropdown choices (ONCE) ===================
+@st.cache_resource(show_spinner=False)
+def _precompute_choices(usa_2024: pd.DataFrame, usa_2025: pd.DataFrame, label: str) -> Dict[str, List]:
+    """
+    Build a dict of column -> list of allowed choices.
+    - Categorical: top-k frequent from union(2024, 2025)
+    - Numeric: percentile grid + most frequent rounded values
+    """
+    CHOICES: Dict[str, List] = {}
+    both = pd.concat([usa_2024, usa_2025.reindex(columns=usa_2024.columns, fill_value=np.nan)], axis=0, ignore_index=True)
+    for col in usa_2024.columns:
+        if col == label:  # never expose label to edit
+            continue
+
+        s = both[col]
+        if pd.api.types.is_numeric_dtype(usa_2024[col]):
+            s_num = pd.to_numeric(s, errors="coerce").dropna()
+            if len(s_num) == 0:
+                CHOICES[col] = [0]
+                continue
+
+            # Percentile grid (5th..95th) + median
+            q_list = [5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 95]
+            qs = np.percentile(s_num, q_list).tolist()
+
+            # Common rounded values (to nearest 1000) by frequency
+            rounded = (np.round(s_num / 1000) * 1000).astype(int)
+            top_round = rounded.value_counts().head(10).index.astype(int).tolist()
+
+            # Merge, deduplicate, sort, and keep as Python numbers (not numpy types)
+            merged = sorted(set(int(v) for v in qs + top_round))
+            CHOICES[col] = merged[:50]  # cap length
+        else:
+            s_cat = s.astype(str).replace({"nan": None})
+            s_cat = s_cat.dropna()
+            if len(s_cat) == 0:
+                CHOICES[col] = [""]
+                continue
+            topk = s_cat.value_counts().head(40).index.tolist()
+            CHOICES[col] = topk
+    return CHOICES
+
+CHOICES_DICT = _precompute_choices(usa_data, usa_25, LABEL)
+
+# =================== Core sampling funcs (your logic) ===================
+def _sample_from_2024(colname: str, df_2024: pd.DataFrame, rng: np.random.Generator):
+    series = df_2024[colname].dropna()
+    if series.empty:
+        if pd.api.types.is_numeric_dtype(df_2024[colname]):
+            return 0
+        return ""
+    return series.sample(1, random_state=rng.integers(0, 10_000)).iloc[0]
+
+def build_synthetic_row_with_trace(
+    usa_25: pd.DataFrame,
+    usa_2024: pd.DataFrame,
+    label: str = "CompTotal",
+    rng: np.random.Generator | None = None
+) -> Tuple[pd.DataFrame, float | None, Dict[str, str], Dict[str, Any]]:
+    if rng is None:
+        rng = _ensure_rng()
+
+    expected_features = [c for c in usa_2024.columns if c != label]
+    row25 = usa_25.sample(1, random_state=rng.integers(0, 10_000)).iloc[0]
+
+    synthetic = {}
+    source_info: Dict[str, str] = {}
+
+    for col in expected_features:
+        use_25_val = False
+        val = None
+
+        if col in row25.index:
+            val = row25[col]
+            if pd.api.types.is_numeric_dtype(usa_2024[col]):
+                val = pd.to_numeric(val, errors="coerce")
+            if not pd.isna(val):
+                use_25_val = True
+
+        if not use_25_val:
+            val = _sample_from_2024(col, usa_2024, rng)
+            source_info[col] = "2024"
+        else:
+            source_info[col] = "2025"
+
+        if pd.api.types.is_numeric_dtype(usa_2024[col]):
+            val = pd.to_numeric(val, errors="coerce")
+            if pd.isna(val):
+                val = _sample_from_2024(col, usa_2024, rng)
+                source_info[col] = "2024"
+        else:
+            if pd.isna(val):
+                val = _sample_from_2024(col, usa_2024, rng)
+                source_info[col] = "2024"
+            if not isinstance(val, str):
+                val = str(val)
+
+        synthetic[col] = val
+
+    X_one = pd.DataFrame([synthetic], columns=expected_features)
+
+    y_true = None
+    if label in row25.index:
+        y_true = pd.to_numeric(row25[label], errors="coerce")
+        if pd.isna(y_true):
+            y_true = None
+
+    used_from_2025 = sum(v == "2025" for v in source_info.values())
+    used_from_2024 = sum(v == "2024" for v in source_info.values())
+    total = len(source_info)
+    report = dict(
+        total_features=total,
+        n_2025=used_from_2025,
+        n_2024=used_from_2024,
+        pct_2025=0.0 if total == 0 else used_from_2025 / total * 100,
+        pct_2024=0.0 if total == 0 else used_from_2024 / total * 100,
+        filled_cols=[c for c, s in source_info.items() if s == "2024"]
+    )
+    return X_one, y_true, source_info, report
+
+def random_sample_row_2024(
+    df_2024: pd.DataFrame,
+    label: str,
+    rng: np.random.Generator
+) -> pd.DataFrame:
+    expected_features = [c for c in df_2024.columns if c != label]
+    sampled = {}
+    for col in expected_features:
+        sampled[col] = _sample_from_2024(col, df_2024, rng)
+    return pd.DataFrame([sampled], columns=expected_features)
+
+# =================== Session state ===================
+if "history" not in st.session_state:
+    st.session_state.history = []  # list of dicts: pred, truth, abs_err, pct_err
+
+if "prepared" not in st.session_state:
+    st.session_state.prepared = None  # dict: X, y_true, info, report, mode
+
+# =================== Sidebar ===================
+st.sidebar.header("Controls")
+mode = st.sidebar.radio(
+    "Choose input mode:",
+    ["Hybrid (Random 2025 + fill from 2024)", "Pure 2024 synthetic"],
+    index=0
+)
+
+if st.sidebar.button("Reload new random data"):
+    rng = new_rng()
+    if mode.startswith("Hybrid"):
+        X_one, y_true, src_info, rep = build_synthetic_row_with_trace(usa_25, usa_data, label=LABEL, rng=rng)
+        st.session_state.prepared = dict(X=X_one, y_true=y_true, info=src_info, report=rep, mode="hybrid")
+    else:
+        X_2024 = random_sample_row_2024(usa_data, LABEL, rng)
+        st.session_state.prepared = dict(X=X_2024, y_true=None, info={}, report={}, mode="pure2024")
+    st.toast("New random input prepared.", icon="✅")
+
+# =================== Explanation ===================
+with st.expander("What is happening here? "):
+    st.markdown(
+        """
+**Goal:** Create one model-ready row and predict **CompTotal(Annual Income)**.
+**dataset** trained on: USA data from 2024 stackoverflow survey.
+            predict on: synthetic/hybrid data mixing 2025 & 2024.
+**Modes:**
+1) **Hybrid** — Pick a random 2025 respondent; for any missing required feature, fill using a value sampled from the 2024 distribution for that column.
+2) **Pure 2024** — Build an entirely synthetic row, sampling every feature from the 2024 distribution.
+
+**Editing:**  
+Below, you can adjust the *current* row via dropdowns. All dropdown choices were **precomputed once at startup** from the union of 2024 & 2025 data (to keep the app fast).  
+- **Categorical**: most frequent categories (+ always includes the current value).  
+- **Numeric**: percentile grid & common rounded values (+ current value).
+
+Click **Submit & Predict** to see predicted vs. true (if available), absolute and percentage errors, and running averages.
+        """
+    )
+
+st.markdown("---")
+
+# =================== Prepare first candidate if needed ===================
+def _prepare_candidate_if_needed():
+    if st.session_state.prepared is None:
+        rng = _ensure_rng()
+        if mode.startswith("Hybrid"):
+            X_one, y_true, src_info, rep = build_synthetic_row_with_trace(usa_25, usa_data, label=LABEL, rng=rng)
+            st.session_state.prepared = dict(X=X_one, y_true=y_true, info=src_info, report=rep, mode="hybrid")
+        else:
+            X_2024 = random_sample_row_2024(usa_data, LABEL, rng)
+            st.session_state.prepared = dict(X=X_2024, y_true=None, info={}, report={}, mode="pure2024")
+
+_prepare_candidate_if_needed()
+
+# =================== Helper: coerce edited values to correct dtype ===================
+def _coerce_value(col: str, val):
+    if pd.api.types.is_numeric_dtype(usa_data[col]):
+        # Accept numbers that might come from selectbox as str/float
+        try:
+            return pd.to_numeric(val)
+        except Exception:
+            return np.nan
+    else:
+        return "" if val is None else str(val)
+
+
+# =================== Predict & History ===================
+left, right = st.columns([1.1, 0.9], gap="large")
+
+with left:
+    st.subheader("Submit a prediction")
+
+    # Preview current candidate
+    if st.session_state.prepared is not None:
+        curr = st.session_state.prepared
+        mode_tag = "Hybrid (2025+2024)" if curr["mode"] == "hybrid" else "Pure 2024"
+        st.caption(f"Next input mode: **{mode_tag}**")
+
+        with st.expander("Show current input row (after your edits)"):
+            st.dataframe(curr["X"].T.rename(columns={0: "value"}))
+
+        if curr["mode"] == "hybrid":
+            rep = curr["report"]
+            st.caption(
+                f"Data completion: **{rep['n_2025']}** from 2025 "
+                f"({rep['pct_2025']:.1f}%); **{rep['n_2024']}** from 2024 "
+                f"({rep['pct_2024']:.1f}%)"
+            )
+
+    submitted = st.button("Submit & Predict", type="primary", use_container_width=True)
+
+    if submitted and st.session_state.prepared is not None:
+        curr = st.session_state.prepared
+        X_one = curr["X"]
+        y_true = curr["y_true"]
+
+        # (Optional) Final dtype alignment just before predict
+        try:
+            # Align numeric dtypes to training schema to be safe
+            for col in X_one.columns:
+                if pd.api.types.is_numeric_dtype(usa_data[col]):
+                    X_one[col] = pd.to_numeric(X_one[col], errors="coerce")
+                else:
+                    X_one[col] = X_one[col].astype(str).fillna("")
+        except Exception:
+            pass
+
+        try:
+            y_pred = float(pipe.predict(X_one)[0])
+        except Exception as e:
+            st.error(f"Prediction failed: {e}")
+            y_pred = None
+
+        if y_pred is not None:
+            st.success(f"**Predicted CompTotal:** {y_pred:,.0f} USD")
+
+            if y_true is not None:
+                st.info(f"**2025 true:** {y_true:,.0f} USD")
+
+                abs_err = abs(y_pred - y_true)
+                pct_err = abs_err / y_true * 100 if y_true != 0 else np.nan
+
+                st.write(f"**Absolute error:** {abs_err:,.0f} USD")
+                st.write(f"**Percentage error:** {pct_err:.2f}%")
+
+                st.session_state.history.append(
+                    dict(pred=y_pred, truth=y_true, abs_err=abs_err, pct_err=pct_err)
+                )
+            else:
+                st.warning("No ground-truth value available for this input (pure 2024 synthetic).")
+
+        # Prepare a new random candidate (and your edits will apply to the new one next round)
+        rng = new_rng()
+        if mode.startswith("Hybrid"):
+            X_one2, y_true2, src_info2, rep2 = build_synthetic_row_with_trace(usa_25, usa_data, label=LABEL, rng=rng)
+            st.session_state.prepared = dict(X=X_one2, y_true=y_true2, info=src_info2, report=rep2, mode="hybrid")
+        else:
+            X_2024b = random_sample_row_2024(usa_data, LABEL, rng)
+            st.session_state.prepared = dict(X=X_2024b, y_true=None, info={}, report={}, mode="pure2024")
+        st.toast("New random input prepared.", icon="✨")
+
+with right:
+    st.subheader("Results history")
+    if len(st.session_state.history) == 0:
+        st.write("No submissions yet.")
+    else:
+        hist_df = pd.DataFrame(st.session_state.history)
+        st.dataframe(
+            hist_df.style.format({"pred": "{:,.0f}", "truth": "{:,.0f}", "abs_err": "{:,.0f}", "pct_err": "{:.2f}"}),
+            use_container_width=True
+        )
+
+        valid = hist_df.dropna(subset=["truth"])
+        if len(valid) > 0:
+            mae = valid["abs_err"].mean()
+            mape = valid["pct_err"].mean()
+            st.metric(label="Mean Absolute Error (USD)", value=f"{mae:,.0f}")
+            st.metric(label="Mean Absolute Percentage Error", value=f"{mape:.2f}%")
+        else:
+            st.write("No entries with ground truth yet.")
+
+st.markdown("---")
+st.caption("Tip: Use the sidebar **Reload new random data** to resample without submitting.")
+# =================== Editable input row UI ===================
+st.subheader("Edit current input (optional)")
+if st.session_state.prepared is not None:
+    curr = st.session_state.prepared
+    X_row = curr["X"].iloc[0].copy()
+
+    # Two tabs for readability
+    tab_cat, tab_num = st.tabs(["Categorical features", "Numeric features"])
+
+    # Build lists
+    cat_cols = [c for c in X_row.index if not pd.api.types.is_numeric_dtype(usa_data[c])]
+    num_cols = [c for c in X_row.index if pd.api.types.is_numeric_dtype(usa_data[c])]
+
+    with tab_cat:
+        st.caption("Pick from common categories (precomputed). Your current value is preselected.")
+        for col in cat_cols:
+            choices = CHOICES_DICT.get(col, [])
+            # ensure current value is present
+            curr_val = "" if pd.isna(X_row[col]) else str(X_row[col])
+            if curr_val not in choices and curr_val != "":
+                choices = [curr_val] + choices
+            sel = st.selectbox(
+                label=col,
+                options=choices if len(choices) > 0 else [""],
+                index=0 if len(choices) == 0 else (choices.index(curr_val) if curr_val in choices else 0),
+                key=f"edit_cat_{col}",
+            )
+            X_row[col] = _coerce_value(col, sel)
+
+    with tab_num:
+        st.caption("Pick typical numeric values (percentiles/rounded) or keep current.")
+        for col in num_cols:
+            choices = CHOICES_DICT.get(col, [])
+            curr_val = X_row[col]
+            # ensure current value is present and cast to int for display if close to int
+            if pd.isna(curr_val):
+                curr_val = choices[0] if len(choices) else 0
+            # make sure current is in choices
+            if len(choices) == 0:
+                choices = [curr_val]
+            elif curr_val not in choices:
+                choices = [curr_val] + choices
+            sel = st.selectbox(
+                label=col,
+                options=choices,
+                index=choices.index(curr_val) if curr_val in choices else 0,
+                key=f"edit_num_{col}",
+            )
+            X_row[col] = _coerce_value(col, sel)
+
+    # Save edits back
+    st.session_state.prepared["X"].iloc[0] = X_row
+
+