Added gitignore, updated README.md

.gitignore

@@ -0,0 +1,18 @@
+__pycache__/
+*.py[cod]
+.pytest_cache/
+.ruff_cache/
+.mypy_cache/
+.venv/
+venv/
+.local/
+.ultralytics/
+outputs/
+models/
+*.webm
+*.mp4
+*.avi
+*.mov
+*.pt
+*.onnx
+*.engine

README.md

@@ -1,6 +1,6 @@
 ---
 title: Tiny Trigger
-emoji: 📉
+emoji: "🔌"
 colorFrom: red
 colorTo: gray
 sdk: gradio
@@ -9,7 +9,116 @@ python_version: '3.13'
 app_file: app.py
 pinned: false
 license: other
-short_description: 'Repository for huggingface build small challenge, '
+short_description: 'Open-vocabulary video automations with YOLOE and llama.cpp'
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Tiny Trigger
+
+Tiny Trigger is a local-first hackathon prototype for open-vocabulary video
+automations. It uses YOLOE to detect user-supplied classes in a video, then
+evaluates small structured rules that can trigger simulated actions or optional
+webhook POSTs.
+
+The LLM path is intentionally constrained: llama.cpp compiles natural language
+into JSON/YAML automation rules. The app validates those rules before evaluating
+them, and the LLM never emits executable code.
+
+## Run Locally
+
+```bash
+pip install -r requirements.txt
+python app.py
+```
+
+Or with Poetry:
+
+```bash
+poetry install
+poetry run python app.py
+```
+
+The detector defaults to `yoloe-26s-seg.pt`, which Ultralytics downloads on
+first use if it is not already cached.
+
+For small or background objects, use a larger model such as `yoloe-26l-seg.pt`,
+set device to `cuda:0`, lower confidence to `0.05`-`0.15`, and raise image size
+to `960` or `1280`.
+
+Run tests with:
+
+```bash
+pip install -r requirements-dev.txt
+python -m pytest
+```
+
+Or:
+
+```bash
+poetry run pytest
+```
+
+## llama.cpp Rule Compiler
+
+Start a llama.cpp server with the small GGUF model:
+
+```bash
+llama-server -hf ggml-org/Qwen3-1.7B-GGUF:Q4_K_M
+```
+
+Then use the default compiler endpoint in the UI:
+
+```text
+http://127.0.0.1:8080/v1
+```
+
+## Rule Shape
+
+Rules can be written directly as YAML or compiled from natural language:
+
+```yaml
+rules:
+  - name: person-near-steering-wheel
+    when:
+      all:
+        - present: {label: person, min_count: 1}
+        - near: {a: person, b: steering wheel, max_distance: 0.45}
+    gate:
+      enabled: true
+      cooldown: {key: turn-on-pc, minutes: 5}
+    then:
+      - type: webhook
+        name: turn on pc
+```
+
+Initial video conditions include presence, count, and near. Gates include enabled
+state and cooldown. Initial actions are simulated events and optional webhooks.
+Home Assistant, live RTSP/webcam monitoring, and trajectory logic are planned
+later increments.
+
+## Local State
+
+Private local settings and runtime memory live under `.local/`, which is ignored
+by git:
+
+```text
+.local/config.yaml       private defaults such as camera/webhook URLs
+.local/automations.json  saved editable automation rules
+.local/state.json        cooldown memory and last-fired timestamps
+.local/events.jsonl      append-only fired action history
+```
+
+Example `.local/config.yaml`:
+
+```yaml
+webhook_url: "http://127.0.0.1:8123/api/webhook/example"
+default_detector_model: "yoloe-26x-seg.pt"
+default_device: "cuda:0"
+default_image_size: 1280
+default_max_detections: 300
+llamacpp_base_url: "http://127.0.0.1:8080/v1"
+```
+
+## License
+
+This repository is provided under a temporary all-rights-reserved hackathon
+prototype license. See `LICENSE`.