diff --git a/internal/demod/gpudemod/README.md b/internal/demod/gpudemod/README.md
index 266c8b8..e51c0ed 100644
--- a/internal/demod/gpudemod/README.md
+++ b/internal/demod/gpudemod/README.md
@@ -21,6 +21,7 @@ This is **not compiled automatically yet** in the current environment because th
 On a CUDA-capable dev machine with toolchain installed:
 
 1. Compile `kernels.cu` into an object file
+   - helper script: `tools/build-gpudemod-kernel.ps1`
 2. Link it into the `cufft` build
 3. Replace `gpud_launch_freq_shift(...)` stub body with the real kernel launch
 4. Validate copied-back shifted IQ against `dsp.FreqShift`
diff --git a/internal/demod/gpudemod/gpudemod.go b/internal/demod/gpudemod/gpudemod.go
index 2b2cd5c..56c2476 100644
--- a/internal/demod/gpudemod/gpudemod.go
+++ b/internal/demod/gpudemod/gpudemod.go
@@ -4,6 +4,7 @@ package gpudemod
 
 /*
 #cgo windows LDFLAGS: -lcufft64_12 -lcudart64_13
+#cgo windows CFLAGS: -I"C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v13.2/include"
 #include <cuda_runtime.h>
 #include <cufft.h>
 
diff --git a/tools/build-gpudemod-kernel.ps1 b/tools/build-gpudemod-kernel.ps1
new file mode 100644
index 0000000..5c8a96b
--- /dev/null
+++ b/tools/build-gpudemod-kernel.ps1
@@ -0,0 +1,27 @@
+param(
+  [string]$CudaRoot = 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.2',
+  [string]$Source = 'internal/demod/gpudemod/kernels.cu',
+  [string]$OutDir = 'internal/demod/gpudemod/build'
+)
+
+$ErrorActionPreference = 'Stop'
+$repo = Split-Path -Parent $PSScriptRoot
+Set-Location $repo
+
+$nvcc = Join-Path $CudaRoot 'bin\nvcc.exe'
+if (!(Test-Path $nvcc)) {
+  throw "nvcc not found at $nvcc"
+}
+
+New-Item -ItemType Directory -Force -Path $OutDir | Out-Null
+$outObj = Join-Path $OutDir 'kernels.obj'
+
+Write-Host "Using nvcc: $nvcc"
+Write-Host "Building $Source -> $outObj"
+
+& $nvcc -c $Source -o $outObj -I (Join-Path $CudaRoot 'include') -Xcompiler "/EHsc"
+if ($LASTEXITCODE -ne 0) {
+  throw "nvcc failed with exit code $LASTEXITCODE"
+}
+
+Write-Host "Built: $outObj"