diff --git a/project_dgx.md b/project_dgx.md
new file mode 100644
index 0000000..d61b1c0
--- /dev/null
+++ b/project_dgx.md
@@ -0,0 +1,25 @@
+---
+name: DGX Spark Nodes
+description: dgxmain and dgxsec - DGX Spark GB10 ARM nodes with full AI stack, Ollama, LiteLLM, credentials
+type: project
+originSessionId: f6f70bf9-8d74-459e-a0c1-f9f2333923d4
+---
+## dgxmain (10.100.102.241, Tailscale 100.124.217.84)
+- ARM Cortex-X925+A725, 20 cores, 122GB RAM, NVIDIA GB10 GPU
+- 3.6TB NVMe (29% used)
+- SSH: yohay / Biton24680#@$ (login), Biton24680#@$ (sudo)
+- Ollama: running natively (systemd), port 11434
+- Models: nemotron-3-super:120b, gpt-oss:120b, llama3:70b, deepseek-r1:70b, qwen3:32b, gemma3:27b, devstral:23.6b, gemma3:4b, nomic-embed-text
+- LiteLLM proxy: port 4000, key sk-dgxmain-litellm-2026
+- Docker containers: Dify stack, Open WebUI, Gitea, n8n, Portainer, Caddy, WordPress, ArangoDB, TensorRT-LLM, Kali
+
+## dgxsec (10.100.102.240, Tailscale 100.78.185.72)
+- Same hardware as dgxmain
+- 3.6TB NVMe (75% used - needs cleanup)
+- SSH: yohay / Biton24680#@$ (login), Biton24680#@$ (sudo)
+- Docker installed but was stopped, now started
+- Ollama: not yet deployed
+
+**Why:** These are the most powerful compute nodes in the cluster. Not in K8s due to ARM compatibility issues.
+
+**How to apply:** Use standalone Docker Compose for GPU workloads. LiteLLM on dgxmain is the unified inference API for all services.