Changes made to the website

README.md

@@ -1,26 +1,51 @@
 ---
 title: SceneSolver
-author: samxkevin
+emoji: 🕵️
+colorFrom: gray
+colorTo: red
 sdk: gradio
-sdk_version: 5.49.1
+sdk_version: "4.44.0"
+python_version: "3.10"
 app_file: app.py
-license: mit
-short_description: Multimodal surveillance analysis and evidence synthesis
+pinned: false
+short_description: Nine-stage forensic video analysis agent (SceneSolver)
+suggested_hardware: zero-a10g
 ---
 
 # SceneSolver
 
-SceneSolver is a multimodal surveillance-analysis system. The public demo exposes the inference pipeline while the trained model assets remain in the author's Hugging Face repositories.
-
-## Pipeline
-
-AE anomaly signal → TimeSformer binary gate → TimeSformer multi-class classification → YOLO weapon detection → audio features → tracking/behavior extraction → fusion reasoning → RL3033 temporal reasoning → optional LLaVA contextual explanation.
-
-## Model sources
-
-- `samxkevin/scenesolver-assets`
-- `samxkevin/SceneDescriptionModel`
-
-## Notes
-
-The system is an experimental/research project. Results are not definitive conclusions and require human review.
+A nine-stage forensic video analysis pipeline (AutoEncoder anomaly signal,
+TimeSformer binary + multi-class classification, YOLO weapon detection, audio
+feature extraction, motion tracking, a fusion reasoning layer, an RL3033
+temporal-reasoning agent, and a LLaVA-based contextual explanation stage),
+wrapped in a Gradio interface with an agentic "Sherl0ck" escalation layer.
+
+Model assets: `samxkevin/scenesolver-assets` (pipeline weights) and
+`samxkevin/SceneDescriptionModel` (the VLM used by Stage 8 / LLaVA).
+
+## Running on ZeroGPU — before you deploy
+
+- **Creating** a ZeroGPU Space requires a Hugging Face **PRO** personal account
+  or a Team/Enterprise organization. (Anyone can *use* a public ZeroGPU Space
+  for free once it exists — the PRO/Enterprise requirement is only for the
+  owner who creates it.)
+- Setting `suggested_hardware: zero-a10g` above is a display hint only. You
+  still need to explicitly select **ZeroGPU** in the Space's
+  **Settings → Hardware** panel for it to actually run there.
+- `COHERE_API_KEY` must be set as a **repository secret**
+  (Settings → Variables and secrets) — it is read from the environment and is
+  no longer hardcoded in the source (see the security note below).
+- The `@spaces.GPU(duration=300)` timeout on the Analyse pipeline is a
+  starting estimate covering AutoEncoder + TimeSformer (x2) + YOLO + audio +
+  tracking + fusion + RL3033 + optional LLaVA. Watch the Space logs after the
+  first real run and raise/lower `duration` in `app.py` to match observed
+  runtime.
+- Verify `samxkevin/scenesolver-assets` and `samxkevin/SceneDescriptionModel`
+  are public repos. If either is private, add an `HF_TOKEN` repository secret
+  with read access, or the Space will fail to download weights at startup.
+
+## Security note
+
+The uploaded source notebook contained a hardcoded, live Cohere API key. It
+has been removed from this port. **That key should be rotated in the Cohere
+dashboard**, since it was exposed in plaintext in the original file.

requirements.txt

@@ -1,20 +1,35 @@
-gradio>=5.0,<6.0
-spaces>=0.38.0
-huggingface_hub>=0.34.0
-numpy
+# SceneSolver — Hugging Face Space (ZeroGPU) requirements
+# Derived directly from the imports actually used in app.py (source: NewChecker.ipynb).
+# NOTE: torch/torchvision versions should be checked against the current
+# ZeroGPU CUDA build before first deploy — see
+# https://huggingface.co/docs/hub/spaces-zerogpu (the backing GPU/CUDA image
+# changes over time, and torch must match the wheel it was built against).
+
+spaces
+gradio>=4.44.0
+
+torch>=2.1.0
+torchvision>=0.16.0
+
+transformers>=4.40.0
+huggingface_hub>=0.23.0
+
 opencv-python-headless
-torch
-torchvision
-transformers>=4.45
-accelerate
-safetensors
+numpy
+matplotlib
 Pillow
 tqdm
-matplotlib
+
 reportlab
+
 stable-baselines3
 ultralytics
 lap
+
 librosa
-openai-whisper
+soundfile
+
 cohere>=5.0.0
+liquid-audio
+
+requests