diff --git a/hpo/README.md b/hpo/README.md
new file mode 100644
index 00000000..73b4ca92
--- /dev/null
+++ b/hpo/README.md
@@ -0,0 +1,29 @@
+# HPO with Modal
+
+## Beware, Here Be Dragons
+
+Running hyperparameter optimization on Modal with ultrathink compute is **extremely expensive**. A single HPO sweep can burn through significant cloud credits in minutes.
+
+### Before You Run Anything
+
+1. **Set budget limits** in your Modal dashboard
+2. **Start with small trial counts** (e.g., 3-5 trials, not 100)
+3. **Use cheaper GPU tiers first** (T4) before scaling to A100s
+4. **Monitor costs in real-time** during runs
+
+### Cost Estimation
+
+| GPU Type | Approx. Cost/Hour | HPO Risk Level |
+|----------|-------------------|----------------|
+| T4       | ~$0.60           | Low            |
+| A10G     | ~$1.10           | Medium         |
+| A100-40G | ~$3.00           | High           |
+| A100-80G | ~$4.50           | Very High      |
+
+A 100-trial HPO sweep with 30-minute trials on A100-80G = **~$225**
+
+### Recommended Workflow
+
+1. Debug locally with CPU/small data
+2. Run 3 trials on T4 to validate setup
+3. Scale up incrementally with budget caps
diff --git a/hpo/__init__.py b/hpo/__init__.py
new file mode 100644
index 00000000..a063f7b2
--- /dev/null
+++ b/hpo/__init__.py
@@ -0,0 +1 @@
+# HPO with Modal integration