Run Gemini outside Streamlit session thread

Use a bounded cached worker and responsive polling so hosted WebSocket sessions remain alive while Gemini processes evidence.

app.py

@@ -1,461 +1,488 @@
-from __future__ import annotations
-
-import io
-import os
-from pathlib import Path
-
-import pandas as pd
-import plotly.express as px
-import streamlit as st
-
-from datapilot.analyst import dataframe_csv, gemini_dataset_summary, inspect_dataset
-from datapilot.config import get_settings
-from datapilot.data import SAMPLE_DATASETS, load_sample
-from datapilot.workflow import run_analysis
-
-st.set_page_config(
-    page_title="DataPilot · Autonomous Data Analyst",
-    page_icon="✦",
-    layout="wide",
-    initial_sidebar_state="expanded",
-)
-
-st.markdown(
-    """
-<style>
-@import url('https://fonts.googleapis.com/css2?family=DM+Sans:wght@400;500;600;700&family=Manrope:wght@600;700;800&display=swap');
-:root{--navy:#07111f;--panel:#0e1b2c;--line:#203149;--cyan:#49d7c5;--blue:#6d8dff;--text:#edf4ff;--muted:#91a1b7}
-.stApp{background:radial-gradient(circle at 75% -10%,#17355c 0,transparent 35%),#07111f;color:var(--text)}
-html,body,[class*="css"]{font-family:"DM Sans",sans-serif}
-h1,h2,h3{font-family:"Manrope",sans-serif;letter-spacing:-.03em}
-header[data-testid="stHeader"]{background:transparent}
-div[data-testid="stSidebar"]{background:#091522;border-right:1px solid var(--line)}
-.block-container{max-width:1480px;padding-top:1.1rem;padding-bottom:4rem}
-.brand{display:flex;gap:.75rem;align-items:center;font:800 1.2rem Manrope;color:white;margin:.2rem 0 1.3rem}
-.brand-mark{display:grid;place-items:center;width:34px;height:34px;border-radius:10px;background:linear-gradient(135deg,var(--cyan),var(--blue));color:#07111f}
-.hero{border:1px solid #29405d;background:linear-gradient(125deg,rgba(17,35,57,.96),rgba(9,22,38,.88));border-radius:24px;padding:2rem 2.2rem;margin-bottom:1rem;overflow:hidden;position:relative}
-.hero:after{content:"";position:absolute;width:340px;height:340px;border-radius:50%;right:-100px;top:-190px;background:rgba(73,215,197,.10)}
-.eyebrow{color:var(--cyan);font-size:.73rem;font-weight:700;letter-spacing:.18em;text-transform:uppercase}
-.hero h1{font-size:clamp(2.1rem,4vw,4rem);line-height:1.02;margin:.45rem 0 .7rem;color:white}
-.hero p{max-width:790px;color:#aebdd0;font-size:1.02rem;line-height:1.65;margin:0}
-.stepbar{display:flex;gap:.5rem;flex-wrap:wrap;margin-top:1.4rem}.step{border:1px solid #2d4664;border-radius:999px;padding:.4rem .72rem;color:#9eafc4;font-size:.75rem}.step.on{color:#07111f;background:var(--cyan);border-color:var(--cyan);font-weight:700}
-.panel{background:rgba(14,27,44,.92);border:1px solid var(--line);border-radius:18px;padding:1.15rem 1.25rem;height:100%}
-.kicker{color:var(--cyan);font-size:.72rem;font-weight:700;text-transform:uppercase;letter-spacing:.12em}.muted{color:var(--muted);font-size:.87rem;line-height:1.55}
-.signature{background:linear-gradient(145deg,#11263b,#0b1828);border:1px solid #29435f;border-radius:17px;padding:1rem;margin-top:1rem}.signature strong{color:white}.signature a{color:var(--cyan);text-decoration:none;font-size:.83rem}
-div[data-testid="stMetric"]{background:#0d1b2c;border:1px solid var(--line);padding:15px 17px;border-radius:15px}div[data-testid="stMetric"] label{color:#91a1b7}div[data-testid="stMetricValue"]{color:white}
-.stButton>button,.stDownloadButton>button{border:0;border-radius:11px;background:linear-gradient(135deg,#49d7c5,#6d8dff);color:#07111f;font-weight:800}
-.stButton>button:hover,.stDownloadButton>button:hover{color:#07111f;filter:brightness(1.08)}
-div[data-testid="stFileUploaderDropzone"]{background:#0c1a2b;border:1.5px dashed #3b617c;border-radius:16px;padding:1.3rem}
-div[data-baseweb="tab-list"]{gap:.3rem;background:#0b1828;border:1px solid var(--line);border-radius:13px;padding:.3rem}
-button[data-baseweb="tab"]{border-radius:9px;color:#9caec3}button[data-baseweb="tab"][aria-selected="true"]{background:#172b41;color:white}
-.stDataFrame{border:1px solid var(--line);border-radius:13px;overflow:hidden}
-[data-testid="stAlert"]{border-radius:13px}
-</style>
-""",
-    unsafe_allow_html=True,
-)
-
-settings = get_settings()
-for key, default in {
-    "frame": None,
-    "dataset_name": "",
-    "profile": None,
-    "result": None,
-    "ai_summary": "",
-    "chat": [],
-    "target": None,
-}.items():
-    if key not in st.session_state:
-        st.session_state[key] = default
-
-
-def read_upload(uploaded) -> pd.DataFrame:
-    suffix = Path(uploaded.name).suffix.lower()
-    raw = uploaded.getvalue()
-    if len(raw) > settings.max_upload_mb * 1_048_576:
-        raise ValueError(f"File exceeds the {settings.max_upload_mb} MB limit.")
-    stream = io.BytesIO(raw)
-    if suffix in {".csv", ".tsv", ".txt"}:
-        return pd.read_csv(stream, sep="\t" if suffix == ".tsv" else None, engine="python")
-    if suffix in {".xlsx", ".xls"}:
-        return pd.read_excel(stream)
-    if suffix == ".parquet":
-        return pd.read_parquet(stream)
-    if suffix == ".json":
-        try:
-            return pd.read_json(stream)
-        except ValueError:
-            stream.seek(0)
-            return pd.read_json(stream, lines=True)
-    raise ValueError("Use CSV, TSV, Excel, JSON, or Parquet.")
-
-
-with st.sidebar:
-    st.markdown(
-        '<div class="brand"><span class="brand-mark">✦</span>DataPilot</div>',
-        unsafe_allow_html=True,
-    )
-    st.caption("AUTONOMOUS ANALYSIS WORKSPACE")
-    st.markdown("##### Gemini intelligence")
-    server_api_key = os.getenv("GEMINI_API_KEY", os.getenv("GOOGLE_API_KEY", ""))
-    user_api_key = st.text_input(
-        "Personal Gemini API key (optional)",
-        value="",
-        type="password",
-        help="Leave blank to use the secured server-side key. Never stored or logged.",
-    )
-    api_key = user_api_key.strip() or server_api_key
-    model = st.selectbox("Model", ["gemini-2.5-flash", "gemini-2.5-pro", "gemini-2.0-flash"])
-    st.caption(
-        "● AI ready · secured server key"
-        if server_api_key
-        else ("● AI ready" if api_key else "○ Local analysis mode")
-    )
-    st.divider()
-    st.markdown("##### Privacy controls")
-    metadata_only = st.toggle(
-        "Metadata-first AI",
-        value=True,
-        help="Send schema, aggregate statistics, and three redacted examples—not the full dataset.",
-    )
-    excluded = st.multiselect(
-        "Exclude columns from AI",
-        list(st.session_state.frame.columns) if st.session_state.frame is not None else [],
-    )
-    st.divider()
-    if st.button("Reset workspace", width="stretch"):
-        for key in ("frame", "profile", "result", "ai_summary", "chat", "target"):
-            st.session_state[key] = (
-                None
-                if key in {"frame", "profile", "result", "target"}
-                else ([] if key == "chat" else "")
-            )
-        st.rerun()
-    st.markdown(
-        """
-    <div class="signature">
-      <div class="kicker">Built & designed by</div>
-      <strong>Dinesh Barri</strong><br>
-      <span class="muted">AI Engineer · Data Scientist</span><br><br>
-      <a href="https://github.com/dineshbarri">GitHub ↗</a>&nbsp;&nbsp;
-      <a href="https://www.linkedin.com/in/dinesh-barri-7654b010b">LinkedIn ↗</a>
-    </div>""",
-        unsafe_allow_html=True,
-    )
-
-loaded = st.session_state.frame is not None
-st.markdown(
-    f"""
-<section class="hero">
-  <div class="eyebrow">Evidence-first autonomous data science</div>
-  <h1>Your data. Explained.<br>Decisions, accelerated.</h1>
-  <p>Upload a dataset and DataPilot immediately inspects its structure, surfaces quality risks,
-  recommends analytical targets, creates interactive evidence, and prepares a leakage-safe
-  machine-learning study—with Gemini available for grounded interpretation.</p>
-  <div class="stepbar">
-    <span class="step {"on" if loaded else ""}">01 · Connect</span>
-    <span class="step {"on" if loaded else ""}">02 · Inspect</span>
-    <span class="step {"on" if st.session_state.ai_summary else ""}">03 · Interpret</span>
-    <span class="step {"on" if st.session_state.result else ""}">04 · Model</span>
-    <span class="step {"on" if st.session_state.result else ""}">05 · Deliver</span>
-  </div>
-</section>""",
-    unsafe_allow_html=True,
-)
-
-if not loaded:
-    left, right = st.columns([1.35, 0.65], gap="large")
-    with left:
-        st.markdown('<div class="kicker">Start a new analysis</div>', unsafe_allow_html=True)
-        st.subheader("Drop in your dataset")
-        uploaded = st.file_uploader(
-            "Upload dataset",
-            type=["csv", "tsv", "txt", "xlsx", "xls", "json", "parquet"],
-            label_visibility="collapsed",
-        )
-        st.caption("CSV · TSV · Excel · JSON · Parquet  |  Raw data remains in this session.")
-        if uploaded:
-            try:
-                with st.status("DataPilot is inspecting your dataset…", expanded=True) as status:
-                    st.write("Validating file structure")
-                    frame = read_upload(uploaded)
-                    st.write("Profiling columns, missingness, cardinality, and target candidates")
-                    profile = inspect_dataset(frame)
-                    st.session_state.frame = frame
-                    st.session_state.profile = profile
-                    st.session_state.dataset_name = uploaded.name
-                    status.update(label="Dataset ready", state="complete")
-                st.rerun()
-            except Exception as exc:
-                st.error(f"Upload could not be processed: {exc}")
-    with right:
-        st.markdown(
-            '<div class="panel"><div class="kicker">Try it instantly</div><h3>Explore a trusted demo</h3><p class="muted">Load a complete classification or regression dataset and see the full analyst workflow.</p></div>',
-            unsafe_allow_html=True,
-        )
-        demo = st.selectbox("Demo dataset", list(SAMPLE_DATASETS))
-        if st.button("Load demo workspace", width="stretch"):
-            frame, target, name = load_sample(SAMPLE_DATASETS[demo])
-            st.session_state.frame, st.session_state.target = frame, target
-            st.session_state.dataset_name = name
-            st.session_state.profile = inspect_dataset(frame)
-            st.rerun()
-    st.stop()
-
-frame: pd.DataFrame = st.session_state.frame
-profile = st.session_state.profile or inspect_dataset(frame)
-brief = profile["brief"]
-
-metrics = st.columns(6)
-metrics[0].metric("Rows", f"{brief.rows:,}")
-metrics[1].metric("Columns", f"{brief.columns:,}")
-metrics[2].metric("Numeric", brief.numeric)
-metrics[3].metric("Categorical", brief.categorical)
-metrics[4].metric("Missing cells", f"{brief.missing_cells:,}")
-metrics[5].metric("Quality score", f"{profile['quality_score']}/100")
-
-overview, quality, explore, ai_tab, model_tab, deliver = st.tabs(
-    ["Overview", "Data quality", "Explore", "AI insights", "Model lab", "Deliver"]
-)
-
-with overview:
-    st.subheader(st.session_state.dataset_name)
-    st.caption(f"Dataset fingerprint {brief.fingerprint} · {brief.memory_mb:.2f} MB in memory")
-    first, last, sample = st.tabs(["First 5 rows", "Last 5 rows", "Random sample"])
-    first.dataframe(frame.head(), width="stretch", hide_index=True)
-    last.dataframe(frame.tail(), width="stretch", hide_index=True)
-    sample.dataframe(
-        frame.sample(min(5, len(frame)), random_state=42), width="stretch", hide_index=True
-    )
-    st.markdown("#### Data dictionary")
-    st.dataframe(
-        profile["dictionary"].drop(columns=["issue_count"]), width="stretch", hide_index=True
-    )
-
-with quality:
-    a, b = st.columns([0.75, 1.25])
-    with a:
-        st.markdown("#### Quality signals")
-        st.metric("Duplicate rows", f"{brief.duplicate_rows:,}")
-        st.metric("Completeness", f"{100 - brief.missing_cells / max(1, frame.size) * 100:.1f}%")
-        flagged = profile["dictionary"].query("issue_count > 0")
-        st.metric("Flagged columns", len(flagged))
-        st.info("DataPilot reports evidence first. No rows or values are changed without approval.")
-    with b:
-        missing = profile["missing"][profile["missing"] > 0].sort_values()
-        if len(missing):
-            fig = px.bar(
-                x=missing.values,
-                y=missing.index,
-                orientation="h",
-                labels={"x": "Missing values", "y": "Column"},
-                title="Missing values by column",
-                color=missing.values,
-                color_continuous_scale=["#49d7c5", "#6d8dff"],
-            )
-            fig.update_layout(
-                template="plotly_dark",
-                paper_bgcolor="#0e1b2c",
-                plot_bgcolor="#0e1b2c",
-                coloraxis_showscale=False,
-            )
-            st.plotly_chart(fig, width="stretch")
-        else:
-            st.success("No missing values detected.")
-        if len(flagged):
-            st.dataframe(flagged.drop(columns=["issue_count"]), width="stretch", hide_index=True)
-
-with explore:
-    numeric = profile["numeric"]
-    if numeric:
-        selected = st.selectbox("Explore a numerical feature", numeric)
-        c1, c2 = st.columns(2)
-        fig = px.histogram(
-            frame,
-            x=selected,
-            marginal="box",
-            title=f"Distribution of {selected}",
-            color_discrete_sequence=["#49d7c5"],
-        )
-        fig.update_layout(template="plotly_dark", paper_bgcolor="#0e1b2c", plot_bgcolor="#0e1b2c")
-        c1.plotly_chart(fig, width="stretch")
-        if not profile["correlation"].empty:
-            heat = px.imshow(
-                profile["correlation"],
-                text_auto=".2f",
-                aspect="auto",
-                color_continuous_scale=["#1a2940", "#49d7c5", "#f4b860"],
-                title="Numeric correlation map",
-            )
-            heat.update_layout(template="plotly_dark", paper_bgcolor="#0e1b2c")
-            c2.plotly_chart(heat, width="stretch")
-        else:
-            c2.info("Add another numerical column to calculate correlations.")
-        st.dataframe(frame[numeric].describe().T, width="stretch")
-    else:
-        st.info("This dataset has no numerical columns. Use the categorical overview below.")
-    categories = profile["categorical"]
-    if categories:
-        selected_cat = st.selectbox("Explore a categorical feature", categories)
-        counts = frame[selected_cat].astype(str).value_counts().head(20).reset_index()
-        fig = px.bar(
-            counts,
-            x="count",
-            y=selected_cat,
-            orientation="h",
-            title=f"Top values · {selected_cat}",
-            color="count",
-            color_continuous_scale=["#49d7c5", "#6d8dff"],
-        )
-        fig.update_layout(
-            template="plotly_dark",
-            paper_bgcolor="#0e1b2c",
-            plot_bgcolor="#0e1b2c",
-            coloraxis_showscale=False,
-        )
-        st.plotly_chart(fig, width="stretch")
-
-with ai_tab:
-    st.markdown("#### Ask Gemini to interpret the computed evidence")
-    st.caption(
-        "AI interpretation based on dataset metadata and limited redacted samples. Verify against source documentation."
-    )
-    if not api_key:
-        st.warning(
-            "Enter a Gemini API key in the sidebar. Deterministic profiling remains fully available without AI."
-        )
-    if st.button("Generate AI analyst brief", disabled=not bool(api_key)):
+from __future__ import annotations
+
+import io
+import os
+import time
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+import pandas as pd
+import plotly.express as px
+import streamlit as st
+
+from datapilot.analyst import dataframe_csv, gemini_dataset_summary, inspect_dataset
+from datapilot.config import get_settings
+from datapilot.data import SAMPLE_DATASETS, load_sample
+from datapilot.workflow import run_analysis
+
+st.set_page_config(
+    page_title="DataPilot · Autonomous Data Analyst",
+    page_icon="✦",
+    layout="wide",
+    initial_sidebar_state="expanded",
+)
+
+st.markdown(
+    """
+<style>
+@import url('https://fonts.googleapis.com/css2?family=DM+Sans:wght@400;500;600;700&family=Manrope:wght@600;700;800&display=swap');
+:root{--navy:#07111f;--panel:#0e1b2c;--line:#203149;--cyan:#49d7c5;--blue:#6d8dff;--text:#edf4ff;--muted:#91a1b7}
+.stApp{background:radial-gradient(circle at 75% -10%,#17355c 0,transparent 35%),#07111f;color:var(--text)}
+html,body,[class*="css"]{font-family:"DM Sans",sans-serif}
+h1,h2,h3{font-family:"Manrope",sans-serif;letter-spacing:-.03em}
+header[data-testid="stHeader"]{background:transparent}
+div[data-testid="stSidebar"]{background:#091522;border-right:1px solid var(--line)}
+.block-container{max-width:1480px;padding-top:1.1rem;padding-bottom:4rem}
+.brand{display:flex;gap:.75rem;align-items:center;font:800 1.2rem Manrope;color:white;margin:.2rem 0 1.3rem}
+.brand-mark{display:grid;place-items:center;width:34px;height:34px;border-radius:10px;background:linear-gradient(135deg,var(--cyan),var(--blue));color:#07111f}
+.hero{border:1px solid #29405d;background:linear-gradient(125deg,rgba(17,35,57,.96),rgba(9,22,38,.88));border-radius:24px;padding:2rem 2.2rem;margin-bottom:1rem;overflow:hidden;position:relative}
+.hero:after{content:"";position:absolute;width:340px;height:340px;border-radius:50%;right:-100px;top:-190px;background:rgba(73,215,197,.10)}
+.eyebrow{color:var(--cyan);font-size:.73rem;font-weight:700;letter-spacing:.18em;text-transform:uppercase}
+.hero h1{font-size:clamp(2.1rem,4vw,4rem);line-height:1.02;margin:.45rem 0 .7rem;color:white}
+.hero p{max-width:790px;color:#aebdd0;font-size:1.02rem;line-height:1.65;margin:0}
+.stepbar{display:flex;gap:.5rem;flex-wrap:wrap;margin-top:1.4rem}.step{border:1px solid #2d4664;border-radius:999px;padding:.4rem .72rem;color:#9eafc4;font-size:.75rem}.step.on{color:#07111f;background:var(--cyan);border-color:var(--cyan);font-weight:700}
+.panel{background:rgba(14,27,44,.92);border:1px solid var(--line);border-radius:18px;padding:1.15rem 1.25rem;height:100%}
+.kicker{color:var(--cyan);font-size:.72rem;font-weight:700;text-transform:uppercase;letter-spacing:.12em}.muted{color:var(--muted);font-size:.87rem;line-height:1.55}
+.signature{background:linear-gradient(145deg,#11263b,#0b1828);border:1px solid #29435f;border-radius:17px;padding:1rem;margin-top:1rem}.signature strong{color:white}.signature a{color:var(--cyan);text-decoration:none;font-size:.83rem}
+div[data-testid="stMetric"]{background:#0d1b2c;border:1px solid var(--line);padding:15px 17px;border-radius:15px}div[data-testid="stMetric"] label{color:#91a1b7}div[data-testid="stMetricValue"]{color:white}
+.stButton>button,.stDownloadButton>button{border:0;border-radius:11px;background:linear-gradient(135deg,#49d7c5,#6d8dff);color:#07111f;font-weight:800}
+.stButton>button:hover,.stDownloadButton>button:hover{color:#07111f;filter:brightness(1.08)}
+div[data-testid="stFileUploaderDropzone"]{background:#0c1a2b;border:1.5px dashed #3b617c;border-radius:16px;padding:1.3rem}
+div[data-baseweb="tab-list"]{gap:.3rem;background:#0b1828;border:1px solid var(--line);border-radius:13px;padding:.3rem}
+button[data-baseweb="tab"]{border-radius:9px;color:#9caec3}button[data-baseweb="tab"][aria-selected="true"]{background:#172b41;color:white}
+.stDataFrame{border:1px solid var(--line);border-radius:13px;overflow:hidden}
+[data-testid="stAlert"]{border-radius:13px}
+</style>
+""",
+    unsafe_allow_html=True,
+)
+
+
+@st.cache_resource
+def ai_executor() -> ThreadPoolExecutor:
+    """Keep slow provider I/O off Streamlit's session-handling thread."""
+    return ThreadPoolExecutor(max_workers=2, thread_name_prefix="datapilot-ai")
+
+settings = get_settings()
+for key, default in {
+    "frame": None,
+    "dataset_name": "",
+    "profile": None,
+    "result": None,
+    "ai_summary": "",
+    "ai_future": None,
+    "chat": [],
+    "target": None,
+}.items():
+    if key not in st.session_state:
+        st.session_state[key] = default
+
+
+def read_upload(uploaded) -> pd.DataFrame:
+    suffix = Path(uploaded.name).suffix.lower()
+    raw = uploaded.getvalue()
+    if len(raw) > settings.max_upload_mb * 1_048_576:
+        raise ValueError(f"File exceeds the {settings.max_upload_mb} MB limit.")
+    stream = io.BytesIO(raw)
+    if suffix in {".csv", ".tsv", ".txt"}:
+        return pd.read_csv(stream, sep="\t" if suffix == ".tsv" else None, engine="python")
+    if suffix in {".xlsx", ".xls"}:
+        return pd.read_excel(stream)
+    if suffix == ".parquet":
+        return pd.read_parquet(stream)
+    if suffix == ".json":
         try:
-            # Keep the widget rerun free of transient status elements. Some hosted
-            # reverse proxies can drop the Streamlit session while a synchronous
-            # external request is paired with an actively updating spinner.
-            summary = gemini_dataset_summary(frame, profile, api_key, model, excluded)
-            st.session_state.ai_summary = summary
+            return pd.read_json(stream)
+        except ValueError:
+            stream.seek(0)
+            return pd.read_json(stream, lines=True)
+    raise ValueError("Use CSV, TSV, Excel, JSON, or Parquet.")
+
+
+with st.sidebar:
+    st.markdown(
+        '<div class="brand"><span class="brand-mark">✦</span>DataPilot</div>',
+        unsafe_allow_html=True,
+    )
+    st.caption("AUTONOMOUS ANALYSIS WORKSPACE")
+    st.markdown("##### Gemini intelligence")
+    server_api_key = os.getenv("GEMINI_API_KEY", os.getenv("GOOGLE_API_KEY", ""))
+    user_api_key = st.text_input(
+        "Personal Gemini API key (optional)",
+        value="",
+        type="password",
+        help="Leave blank to use the secured server-side key. Never stored or logged.",
+    )
+    api_key = user_api_key.strip() or server_api_key
+    model = st.selectbox("Model", ["gemini-2.5-flash", "gemini-2.5-pro", "gemini-2.0-flash"])
+    st.caption(
+        "● AI ready · secured server key"
+        if server_api_key
+        else ("● AI ready" if api_key else "○ Local analysis mode")
+    )
+    st.divider()
+    st.markdown("##### Privacy controls")
+    metadata_only = st.toggle(
+        "Metadata-first AI",
+        value=True,
+        help="Send schema, aggregate statistics, and three redacted examples—not the full dataset.",
+    )
+    excluded = st.multiselect(
+        "Exclude columns from AI",
+        list(st.session_state.frame.columns) if st.session_state.frame is not None else [],
+    )
+    st.divider()
+    if st.button("Reset workspace", width="stretch"):
+        for key in ("frame", "profile", "result", "ai_summary", "chat", "target"):
+            st.session_state[key] = (
+                None
+                if key in {"frame", "profile", "result", "target"}
+                else ([] if key == "chat" else "")
+            )
+        st.rerun()
+    st.markdown(
+        """
+    <div class="signature">
+      <div class="kicker">Built & designed by</div>
+      <strong>Dinesh Barri</strong><br>
+      <span class="muted">AI Engineer · Data Scientist</span><br><br>
+      <a href="https://github.com/dineshbarri">GitHub ↗</a>&nbsp;&nbsp;
+      <a href="https://www.linkedin.com/in/dinesh-barri-7654b010b">LinkedIn ↗</a>
+    </div>""",
+        unsafe_allow_html=True,
+    )
+
+loaded = st.session_state.frame is not None
+st.markdown(
+    f"""
+<section class="hero">
+  <div class="eyebrow">Evidence-first autonomous data science</div>
+  <h1>Your data. Explained.<br>Decisions, accelerated.</h1>
+  <p>Upload a dataset and DataPilot immediately inspects its structure, surfaces quality risks,
+  recommends analytical targets, creates interactive evidence, and prepares a leakage-safe
+  machine-learning study—with Gemini available for grounded interpretation.</p>
+  <div class="stepbar">
+    <span class="step {"on" if loaded else ""}">01 · Connect</span>
+    <span class="step {"on" if loaded else ""}">02 · Inspect</span>
+    <span class="step {"on" if st.session_state.ai_summary else ""}">03 · Interpret</span>
+    <span class="step {"on" if st.session_state.result else ""}">04 · Model</span>
+    <span class="step {"on" if st.session_state.result else ""}">05 · Deliver</span>
+  </div>
+</section>""",
+    unsafe_allow_html=True,
+)
+
+if not loaded:
+    left, right = st.columns([1.35, 0.65], gap="large")
+    with left:
+        st.markdown('<div class="kicker">Start a new analysis</div>', unsafe_allow_html=True)
+        st.subheader("Drop in your dataset")
+        uploaded = st.file_uploader(
+            "Upload dataset",
+            type=["csv", "tsv", "txt", "xlsx", "xls", "json", "parquet"],
+            label_visibility="collapsed",
+        )
+        st.caption("CSV · TSV · Excel · JSON · Parquet  |  Raw data remains in this session.")
+        if uploaded:
+            try:
+                with st.status("DataPilot is inspecting your dataset…", expanded=True) as status:
+                    st.write("Validating file structure")
+                    frame = read_upload(uploaded)
+                    st.write("Profiling columns, missingness, cardinality, and target candidates")
+                    profile = inspect_dataset(frame)
+                    st.session_state.frame = frame
+                    st.session_state.profile = profile
+                    st.session_state.dataset_name = uploaded.name
+                    status.update(label="Dataset ready", state="complete")
+                st.rerun()
+            except Exception as exc:
+                st.error(f"Upload could not be processed: {exc}")
+    with right:
+        st.markdown(
+            '<div class="panel"><div class="kicker">Try it instantly</div><h3>Explore a trusted demo</h3><p class="muted">Load a complete classification or regression dataset and see the full analyst workflow.</p></div>',
+            unsafe_allow_html=True,
+        )
+        demo = st.selectbox("Demo dataset", list(SAMPLE_DATASETS))
+        if st.button("Load demo workspace", width="stretch"):
+            frame, target, name = load_sample(SAMPLE_DATASETS[demo])
+            st.session_state.frame, st.session_state.target = frame, target
+            st.session_state.dataset_name = name
+            st.session_state.profile = inspect_dataset(frame)
+            st.rerun()
+    st.stop()
+
+frame: pd.DataFrame = st.session_state.frame
+profile = st.session_state.profile or inspect_dataset(frame)
+brief = profile["brief"]
+
+metrics = st.columns(6)
+metrics[0].metric("Rows", f"{brief.rows:,}")
+metrics[1].metric("Columns", f"{brief.columns:,}")
+metrics[2].metric("Numeric", brief.numeric)
+metrics[3].metric("Categorical", brief.categorical)
+metrics[4].metric("Missing cells", f"{brief.missing_cells:,}")
+metrics[5].metric("Quality score", f"{profile['quality_score']}/100")
+
+overview, quality, explore, ai_tab, model_tab, deliver = st.tabs(
+    ["Overview", "Data quality", "Explore", "AI insights", "Model lab", "Deliver"]
+)
+
+with overview:
+    st.subheader(st.session_state.dataset_name)
+    st.caption(f"Dataset fingerprint {brief.fingerprint} · {brief.memory_mb:.2f} MB in memory")
+    first, last, sample = st.tabs(["First 5 rows", "Last 5 rows", "Random sample"])
+    first.dataframe(frame.head(), width="stretch", hide_index=True)
+    last.dataframe(frame.tail(), width="stretch", hide_index=True)
+    sample.dataframe(
+        frame.sample(min(5, len(frame)), random_state=42), width="stretch", hide_index=True
+    )
+    st.markdown("#### Data dictionary")
+    st.dataframe(
+        profile["dictionary"].drop(columns=["issue_count"]), width="stretch", hide_index=True
+    )
+
+with quality:
+    a, b = st.columns([0.75, 1.25])
+    with a:
+        st.markdown("#### Quality signals")
+        st.metric("Duplicate rows", f"{brief.duplicate_rows:,}")
+        st.metric("Completeness", f"{100 - brief.missing_cells / max(1, frame.size) * 100:.1f}%")
+        flagged = profile["dictionary"].query("issue_count > 0")
+        st.metric("Flagged columns", len(flagged))
+        st.info("DataPilot reports evidence first. No rows or values are changed without approval.")
+    with b:
+        missing = profile["missing"][profile["missing"] > 0].sort_values()
+        if len(missing):
+            fig = px.bar(
+                x=missing.values,
+                y=missing.index,
+                orientation="h",
+                labels={"x": "Missing values", "y": "Column"},
+                title="Missing values by column",
+                color=missing.values,
+                color_continuous_scale=["#49d7c5", "#6d8dff"],
+            )
+            fig.update_layout(
+                template="plotly_dark",
+                paper_bgcolor="#0e1b2c",
+                plot_bgcolor="#0e1b2c",
+                coloraxis_showscale=False,
+            )
+            st.plotly_chart(fig, width="stretch")
+        else:
+            st.success("No missing values detected.")
+        if len(flagged):
+            st.dataframe(flagged.drop(columns=["issue_count"]), width="stretch", hide_index=True)
+
+with explore:
+    numeric = profile["numeric"]
+    if numeric:
+        selected = st.selectbox("Explore a numerical feature", numeric)
+        c1, c2 = st.columns(2)
+        fig = px.histogram(
+            frame,
+            x=selected,
+            marginal="box",
+            title=f"Distribution of {selected}",
+            color_discrete_sequence=["#49d7c5"],
+        )
+        fig.update_layout(template="plotly_dark", paper_bgcolor="#0e1b2c", plot_bgcolor="#0e1b2c")
+        c1.plotly_chart(fig, width="stretch")
+        if not profile["correlation"].empty:
+            heat = px.imshow(
+                profile["correlation"],
+                text_auto=".2f",
+                aspect="auto",
+                color_continuous_scale=["#1a2940", "#49d7c5", "#f4b860"],
+                title="Numeric correlation map",
+            )
+            heat.update_layout(template="plotly_dark", paper_bgcolor="#0e1b2c")
+            c2.plotly_chart(heat, width="stretch")
+        else:
+            c2.info("Add another numerical column to calculate correlations.")
+        st.dataframe(frame[numeric].describe().T, width="stretch")
+    else:
+        st.info("This dataset has no numerical columns. Use the categorical overview below.")
+    categories = profile["categorical"]
+    if categories:
+        selected_cat = st.selectbox("Explore a categorical feature", categories)
+        counts = frame[selected_cat].astype(str).value_counts().head(20).reset_index()
+        fig = px.bar(
+            counts,
+            x="count",
+            y=selected_cat,
+            orientation="h",
+            title=f"Top values · {selected_cat}",
+            color="count",
+            color_continuous_scale=["#49d7c5", "#6d8dff"],
+        )
+        fig.update_layout(
+            template="plotly_dark",
+            paper_bgcolor="#0e1b2c",
+            plot_bgcolor="#0e1b2c",
+            coloraxis_showscale=False,
+        )
+        st.plotly_chart(fig, width="stretch")
+
+with ai_tab:
+    st.markdown("#### Ask Gemini to interpret the computed evidence")
+    st.caption(
+        "AI interpretation based on dataset metadata and limited redacted samples. Verify against source documentation."
+    )
+    if not api_key:
+        st.warning(
+            "Enter a Gemini API key in the sidebar. Deterministic profiling remains fully available without AI."
+        )
+    ai_future = st.session_state.ai_future
+    if st.button(
+        "Generate AI analyst brief",
+        disabled=not bool(api_key) or ai_future is not None,
+    ):
+        st.session_state.ai_future = ai_executor().submit(
+            gemini_dataset_summary,
+            frame.copy(deep=True),
+            profile,
+            api_key,
+            model,
+            list(excluded),
+        )
+        st.rerun()
+
+    ai_future = st.session_state.ai_future
+    if ai_future is not None and ai_future.done():
+        try:
+            st.session_state.ai_summary = ai_future.result()
             st.success("AI analyst brief ready")
-        except ValueError as exc:
-            st.warning(str(exc))
-        except Exception:
-            st.error(
-                "AI Insights encountered an unexpected problem. "
-                "Your dataset and deterministic analysis remain available."
-            )
-    if st.session_state.ai_summary:
-        st.markdown(st.session_state.ai_summary)
-    with st.expander("What may be sent to Gemini"):
-        st.write(
-            "Column metadata, aggregate statistics, target candidates, quality score, and up to three redacted example rows."
-        )
-        st.write(
-            "Automatically excluded potential PII:",
-            [
-                c
-                for c in frame.columns
-                if any(
-                    k in str(c).lower() for k in ("email", "phone", "address", "name", "account")
-                )
-            ]
-            or "None detected",
-        )
-
-with model_tab:
-    st.markdown("#### Confirm the analytical target")
-    candidates = pd.DataFrame(profile["targets"])
-    st.dataframe(candidates, width="stretch", hide_index=True)
-    default_target = st.session_state.target or (
-        profile["targets"][0]["column"] if profile["targets"] else frame.columns[-1]
-    )
-    target = st.selectbox(
-        "Target column", list(frame.columns), index=list(frame.columns).index(default_target)
-    )
-    st.caption("DataPilot will not train supervised models until you confirm this selection.")
-    if frame[target].nunique(dropna=True) < 2:
-        st.error("The selected target has fewer than two observed values.")
-    run = st.button(
-        "Run autonomous model study",
-        type="primary",
-        disabled=frame[target].nunique(dropna=True) < 2,
-    )
-    if run:
-        try:
-            progress = st.progress(0, text="Preparing agent graph")
-            progress.progress(12, text="Data Quality Agent · auditing risks")
-            with st.spinner(
-                "LangGraph agents are profiling, planning, training, evaluating, and explaining…"
-            ):
-                result = run_analysis(frame, target, st.session_state.dataset_name, settings)
-            progress.progress(100, text="Analysis complete")
-            st.session_state.result = result.model_dump(mode="json")
-            st.success(
-                "Model study completed with leakage-safe preprocessing and cross-validation."
-            )
-        except Exception as exc:
-            st.error(f"Model study failed: {exc}")
-    result = st.session_state.result
-    if result:
-        best = result["model_results"][0]
-        c1, c2, c3 = st.columns(3)
-        c1.metric("Selected model", result["best_model"])
-        c2.metric(
-            "One-time test " + best["primary_metric"].replace("_", " ").title(),
-            f"{best['final_test_score']:.3f}",
-        )
-        c3.metric("CV mean", f"{best['cross_validation_mean']:.3f}")
-        results = pd.DataFrame(result["model_results"])
-        fig = px.bar(
-            results.sort_values("selection_score"),
-            x="selection_score",
-            y="name",
-            orientation="h",
-            color="selection_score",
-            title="Training-CV model selection",
-            color_continuous_scale=["#344b69", "#49d7c5"],
-        )
-        fig.update_layout(
-            template="plotly_dark",
-            paper_bgcolor="#0e1b2c",
-            plot_bgcolor="#0e1b2c",
-            coloraxis_showscale=False,
-        )
-        st.plotly_chart(fig, width="stretch")
-        st.dataframe(results, width="stretch", hide_index=True)
-        st.markdown("#### Agent execution trace")
-        st.dataframe(pd.DataFrame(result["trace"]), width="stretch", hide_index=True)
-
-with deliver:
-    st.markdown("#### Export your evidence")
-    c1, c2 = st.columns(2)
-    c1.download_button(
-        "Download original dataset · CSV",
-        dataframe_csv(frame),
-        file_name=f"{Path(st.session_state.dataset_name).stem}_datapilot.csv",
-        mime="text/csv",
-        width="stretch",
-    )
-    c2.download_button(
-        "Download data dictionary · CSV",
-        dataframe_csv(profile["dictionary"].drop(columns=["issue_count"])),
-        file_name="datapilot_data_dictionary.csv",
-        mime="text/csv",
-        width="stretch",
-    )
-    result = st.session_state.result
-    if result:
-        st.markdown("#### Model and report artifacts")
-        columns = st.columns(min(4, len(result["artifacts"])))
-        for column, (name, raw_path) in zip(columns, result["artifacts"].items(), strict=False):
-            path = Path(raw_path)
-            if path.exists():
-                column.download_button(
-                    name.replace("_", " ").title(),
-                    path.read_bytes(),
-                    file_name=path.name,
-                    width="stretch",
-                )
-    else:
-        st.info(
-            "Run a model study to unlock the fitted pipeline, model card, metrics, and HTML report."
-        )
-
-st.caption(
-    "DataPilot provides exploratory decision support. Predictive associations do not establish causality."
-)
+        except ValueError as exc:
+            st.warning(str(exc))
+        except Exception:
+            st.error(
+                "AI Insights encountered an unexpected problem. "
+                "Your dataset and deterministic analysis remain available."
+            )
+        finally:
+            st.session_state.ai_future = None
+    elif ai_future is not None:
+        st.info("Gemini is reviewing the bounded evidence package…")
+        time.sleep(0.5)
+        st.rerun()
+    if st.session_state.ai_summary:
+        st.markdown(st.session_state.ai_summary)
+    with st.expander("What may be sent to Gemini"):
+        st.write(
+            "Column metadata, aggregate statistics, target candidates, quality score, and up to three redacted example rows."
+        )
+        st.write(
+            "Automatically excluded potential PII:",
+            [
+                c
+                for c in frame.columns
+                if any(
+                    k in str(c).lower() for k in ("email", "phone", "address", "name", "account")
+                )
+            ]
+            or "None detected",
+        )
+
+with model_tab:
+    st.markdown("#### Confirm the analytical target")
+    candidates = pd.DataFrame(profile["targets"])
+    st.dataframe(candidates, width="stretch", hide_index=True)
+    default_target = st.session_state.target or (
+        profile["targets"][0]["column"] if profile["targets"] else frame.columns[-1]
+    )
+    target = st.selectbox(
+        "Target column", list(frame.columns), index=list(frame.columns).index(default_target)
+    )
+    st.caption("DataPilot will not train supervised models until you confirm this selection.")
+    if frame[target].nunique(dropna=True) < 2:
+        st.error("The selected target has fewer than two observed values.")
+    run = st.button(
+        "Run autonomous model study",
+        type="primary",
+        disabled=frame[target].nunique(dropna=True) < 2,
+    )
+    if run:
+        try:
+            progress = st.progress(0, text="Preparing agent graph")
+            progress.progress(12, text="Data Quality Agent · auditing risks")
+            with st.spinner(
+                "LangGraph agents are profiling, planning, training, evaluating, and explaining…"
+            ):
+                result = run_analysis(frame, target, st.session_state.dataset_name, settings)
+            progress.progress(100, text="Analysis complete")
+            st.session_state.result = result.model_dump(mode="json")
+            st.success(
+                "Model study completed with leakage-safe preprocessing and cross-validation."
+            )
+        except Exception as exc:
+            st.error(f"Model study failed: {exc}")
+    result = st.session_state.result
+    if result:
+        best = result["model_results"][0]
+        c1, c2, c3 = st.columns(3)
+        c1.metric("Selected model", result["best_model"])
+        c2.metric(
+            "One-time test " + best["primary_metric"].replace("_", " ").title(),
+            f"{best['final_test_score']:.3f}",
+        )
+        c3.metric("CV mean", f"{best['cross_validation_mean']:.3f}")
+        results = pd.DataFrame(result["model_results"])
+        fig = px.bar(
+            results.sort_values("selection_score"),
+            x="selection_score",
+            y="name",
+            orientation="h",
+            color="selection_score",
+            title="Training-CV model selection",
+            color_continuous_scale=["#344b69", "#49d7c5"],
+        )
+        fig.update_layout(
+            template="plotly_dark",
+            paper_bgcolor="#0e1b2c",
+            plot_bgcolor="#0e1b2c",
+            coloraxis_showscale=False,
+        )
+        st.plotly_chart(fig, width="stretch")
+        st.dataframe(results, width="stretch", hide_index=True)
+        st.markdown("#### Agent execution trace")
+        st.dataframe(pd.DataFrame(result["trace"]), width="stretch", hide_index=True)
+
+with deliver:
+    st.markdown("#### Export your evidence")
+    c1, c2 = st.columns(2)
+    c1.download_button(
+        "Download original dataset · CSV",
+        dataframe_csv(frame),
+        file_name=f"{Path(st.session_state.dataset_name).stem}_datapilot.csv",
+        mime="text/csv",
+        width="stretch",
+    )
+    c2.download_button(
+        "Download data dictionary · CSV",
+        dataframe_csv(profile["dictionary"].drop(columns=["issue_count"])),
+        file_name="datapilot_data_dictionary.csv",
+        mime="text/csv",
+        width="stretch",
+    )
+    result = st.session_state.result
+    if result:
+        st.markdown("#### Model and report artifacts")
+        columns = st.columns(min(4, len(result["artifacts"])))
+        for column, (name, raw_path) in zip(columns, result["artifacts"].items(), strict=False):
+            path = Path(raw_path)
+            if path.exists():
+                column.download_button(
+                    name.replace("_", " ").title(),
+                    path.read_bytes(),
+                    file_name=path.name,
+                    width="stretch",
+                )
+    else:
+        st.info(
+            "Run a model study to unlock the fitted pipeline, model card, metrics, and HTML report."
+        )
+
+st.caption(
+    "DataPilot provides exploratory decision support. Predictive associations do not establish causality."
+)