Polish README for public repo

README.md

@@ -13,6 +13,21 @@ short_description: "[WIP] Interactive visualization of an LLM training cluster"
 
 Interactive workbench for exploring how large-model training layouts map onto GPU clusters.
 
+Live demo: https://huggingface.co/spaces/joeddav/illustrated-cluster
+
+This project is meant to make training-parallelism tradeoffs legible:
+
+- per-GPU memory pressure
+- tensor / pipeline / context / expert communication
+- pipeline bubbles and throughput estimates
+- physical placement across nodes and racks
+
+Status:
+
+- estimates are directional, not production-grade
+- the app is still a WIP and may contain bugs or logical errors
+- the Llama 3.1 405B example is temporarily hidden while its training recipe is being reworked
+
 Current WIP scope:
 
 - compute-backed memory, communication, and throughput estimates
@@ -20,10 +35,6 @@ Current WIP scope:
 - editable model, cluster, training, and parallelism controls
 - built-in OLMo 3 32B and Trinity Large 400B starting points
 
-Temporary note:
-
-- the Llama 3.1 405B example is hidden from the UI while its training recipe is being reworked
-
 ## Stack
 
 - React 19 + TypeScript