Remove development status from README

README.md

@@ -119,35 +119,3 @@ or granting the app write access.
 Do not enable remote writes for arbitrary uploads to a public dataset. Processed
 slides can contain the original lecture material. Use a private dataset for personal
 testing or only prewarm material you have permission to publish.
-
-## Status
-**Professor agent loop working:** upload a lecture PDF → render and index the complete
-deck → Nemotron plans short teaching beats → the orchestrator executes validated slide
-navigation and whiteboard tools → VoxCPM speaks the explanation. Student questions
-cancel active narration, can send the professor to a more relevant slide, and can add
-supporting notes or equations to the whiteboard.
-
-Targeted vision via `look_closer` and richer Excalidraw diagrams remain follow-up work.
-
-Layout: [`app.py`](app.py) (Gradio UI) · [`ai_prof/pdf_utils.py`](ai_prof/pdf_utils.py) (PDF→slides) ·
-[`ai_prof/vision.py`](ai_prof/vision.py) (eyes) · [`ai_prof/brain.py`](ai_prof/brain.py) (brain) ·
-[`ai_prof/config.py`](ai_prof/config.py) (endpoints / mock fallback).
-
-The inference services run on Modal: MiniCPM-V through llama.cpp, Nemotron
-through vLLM, VoxCPM2 through vLLM-Omni, and distil-Whisper through a small
-OpenAI-compatible FastAPI server.
-The Gradio frontend remains deployable as a Hugging Face Space.
-
-Run the focused test suite with:
-
-```bash
-.venv/bin/python -m unittest discover -s tests -v
-```
-
-Next up (see [IDEATION.md](IDEATION.md)): prepared-deck storage on Hugging Face,
-animated structured diagrams, stronger cancellation during browser audio playback,
-and targeted `look_closer` vision.
-
-The current target design is documented in [ARCHITECTURE.md](ARCHITECTURE.md), including complete-deck
-indexing, the professor tool loop, a preprocessed demo lecture, synchronized speech and whiteboard actions,
-and interruption/resume behavior.

app.py

@@ -588,6 +588,30 @@ _BANNER = (
 )
 
 _CSS = """
+.gradio-container {
+    --app-text: #24283b;
+    --app-muted: #667085;
+    --app-card: #ffffff;
+    --app-border: #e1e5eb;
+    --app-panel-title: #ecebff;
+    --app-panel-title-border: #dedcff;
+    --app-accent-text: #5b55c7;
+    --app-slide-bg: #f8f9fb;
+    --app-nav-bg: #f7f6ff;
+    --app-nav-border: #d8d6ff;
+}
+.dark .gradio-container {
+    --app-text: #f2f4f7;
+    --app-muted: #a7b0c0;
+    --app-card: #171923;
+    --app-border: #303442;
+    --app-panel-title: #24213d;
+    --app-panel-title-border: #3a3560;
+    --app-accent-text: #c6c2ff;
+    --app-slide-bg: #10121a;
+    --app-nav-bg: #27243f;
+    --app-nav-border: #4b4678;
+}
 .gradio-container {
     max-width: 1320px !important;
     margin: 0 auto !important;
@@ -600,14 +624,14 @@ _CSS = """
 }
 .app-title h1 {
     margin: 0 0 4px !important;
-    color: #24283b;
+    color: var(--app-text);
     font-size: 2.15rem !important;
     font-weight: 800 !important;
     letter-spacing: -.035em;
 }
 .app-title p {
     margin: 0 !important;
-    color: #697386;
+    color: var(--app-muted);
     font-size: 1rem;
 }
 .workspace-row {
@@ -620,9 +644,9 @@ _CSS = """
     min-width: 0 !important;
     overflow: hidden;
     gap: 0 !important;
-    border: 1px solid #e1e5eb;
+    border: 1px solid var(--app-border);
     border-radius: 14px;
-    background: #fff;
+    background: var(--app-card);
 }
 .panel-title {
     flex: 0 0 46px !important;
@@ -632,9 +656,9 @@ _CSS = """
     align-items: center !important;
     margin: 0 !important;
     padding: 0 14px !important;
-    background: #ecebff;
-    border-bottom: 1px solid #dedcff;
-    color: #5b55c7;
+    background: var(--app-panel-title);
+    border-bottom: 1px solid var(--app-panel-title-border);
+    color: var(--app-accent-text);
 }
 .panel-title p {
     margin: 0 !important;
@@ -656,7 +680,7 @@ _CSS = """
 .slide-frame img {
     height: 550px !important;
     object-fit: contain !important;
-    background: #f8f9fb;
+    background: var(--app-slide-bg);
 }
 .slide-footer {
     padding: 0 12px 12px !important;
@@ -664,7 +688,7 @@ _CSS = """
 .slide-caption {
     min-height: 24px;
     margin: 0 !important;
-    color: #667085;
+    color: var(--app-muted);
 }
 .slide-index {
     margin: 2px 0 8px !important;
@@ -678,9 +702,13 @@ _CSS = """
     font-weight: 700 !important;
 }
 .nav-button button {
-    color: #5b55c7 !important;
-    border: 1px solid #d8d6ff !important;
-    background: #f7f6ff !important;
+    color: var(--app-accent-text) !important;
+    border: 1px solid var(--app-nav-border) !important;
+    background: var(--app-nav-bg) !important;
+}
+.nav-button button:hover {
+    border-color: #7770ef !important;
+    background: color-mix(in srgb, var(--app-nav-bg) 82%, #625ce7) !important;
 }
 .explain-button button {
     color: #fff !important;
@@ -795,7 +823,7 @@ _CSS = """
 }
 .mic-label {
     margin: 2px 0 -4px !important;
-    color: #697386;
+    color: var(--app-muted);
     font-size: .82rem;
     font-weight: 650;
 }
@@ -828,9 +856,18 @@ _CSS = """
 }
 .upload-copy {
     margin: 0 !important;
-    color: #667085;
+    color: var(--app-muted);
     font-size: .88rem;
 }
+.dark .panel-card input,
+.dark .panel-card textarea,
+.dark .panel-card select {
+    color-scheme: dark;
+}
+.dark .status-strip > div[style] {
+    filter: brightness(.72) saturate(.9);
+    color: #f3f4f6 !important;
+}
 @media (max-width: 900px) {
     .gradio-container { padding-inline: 12px !important; }
     .teaching-panel, .bottom-panel { min-width: 100% !important; }
@@ -883,7 +920,7 @@ with gr.Blocks(title="AI Prof", theme=gr.themes.Soft(), css=_CSS) as demo:
     with gr.Row(equal_height=True, elem_classes=["workspace-row"]):
         with gr.Column(scale=5, elem_classes=["panel-card", "bottom-panel"]):
             gr.Markdown("Lecture transcript", elem_classes=["panel-title"])
-            status_strip = gr.HTML(value=_STATUS_IDLE)
+            status_strip = gr.HTML(value=_STATUS_IDLE, elem_classes=["status-strip"])
             prof_audio = gr.Audio(
                 autoplay=True,
                 show_label=False,

modal_app.py

@@ -72,6 +72,10 @@ def _vllm_cmd() -> list[str]:
         "--max-model-len", str(MAX_MODEL_LEN),
         "--max-num-seqs", "8",
         "--tensor-parallel-size", "1",
+        # Modal Volumes mount as 9P, which vLLM does not recognize as a network
+        # filesystem. Force parallel prefetch instead of reading 13 shards
+        # serially; the serial path takes roughly ten minutes for this checkpoint.
+        "--safetensors-load-strategy", "prefetch",
         "--trust-remote-code",
         "--reasoning-parser", "nemotron_v3",
     ]
@@ -132,6 +136,7 @@ def warm() -> None:
             proc.kill()
     vllm_cache.commit()
     flashinfer_cache.commit()
+    triton_cache.commit()
     print("Warm complete — compile caches committed. Cold starts will now be fast.")