diff --git a/deploy/parakeet-stt/edge-stt/README.md b/deploy/parakeet-stt/edge-stt/README.md
index 935f276..b66783f 100644
--- a/deploy/parakeet-stt/edge-stt/README.md
+++ b/deploy/parakeet-stt/edge-stt/README.md
@@ -47,7 +47,9 @@ Used **as-is** on a **shared** Jetson — no reflash, no clock lock (see [`ASSUM
 source scripts/preflight.sh           # or: ./transcribe --selfcheck
 
 # 0b. Install the locked Python deps (scoring + data-prep; pinned for aarch64/cp310).
-pip install --break-system-packages -r requirements.lock
+#     NOTE: PyPI can be slow on a Jetson (~tens of kB/s seen) — the ~17 MB of wheels may take
+#     several minutes; add `-v` if you want progress. It is downloading, not hung.
+pip install -v --break-system-packages -r requirements.lock
 
 # 1. Eval data: LibriSpeech test-clean first-300-by-id -> data/wav/ + data/manifest.jsonl
 bash scripts/get-data.sh
diff --git a/deploy/parakeet-stt/edge-stt/scripts/build-parakeet.sh b/deploy/parakeet-stt/edge-stt/scripts/build-parakeet.sh
index caae11d..0a243c0 100644
--- a/deploy/parakeet-stt/edge-stt/scripts/build-parakeet.sh
+++ b/deploy/parakeet-stt/edge-stt/scripts/build-parakeet.sh
@@ -16,7 +16,8 @@ set -euo pipefail
 
 SRC="${PARAKEET_SRC:-$HOME/parakeet.cpp}"
 TAG="${PARAKEET_TAG:-v0.1.1}"
-JOBS="${JOBS:-4}"   # 8 GB unified RAM is tight; bound parallelism
+JOBS="${JOBS:-2}"   # 8 GB unified RAM is tight: the ggml-cuda nvcc compile spikes ~1-2 GB/job, so
+                    # -j4 can OOM-reboot the box (observed). -j2 is the safe default; override with JOBS=.
 
 if [ ! -d "$SRC/.git" ]; then
   echo "[build-parakeet] cloning $TAG"