diff --git a/build-cuda-linux.sh b/build-cuda-linux.sh
new file mode 100644
index 0000000..f94ec08
--- /dev/null
+++ b/build-cuda-linux.sh
@@ -0,0 +1,18 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+CUDA_ROOT="${CUDA_ROOT:-/usr/local/cuda}"
+SRC="internal/demod/gpudemod/kernels.cu"
+OUT_DIR="internal/demod/gpudemod/build"
+OUT_OBJ="$OUT_DIR/kernels.o"
+
+mkdir -p "$OUT_DIR"
+
+if [[ ! -x "$CUDA_ROOT/bin/nvcc" ]]; then
+  echo "nvcc not found at $CUDA_ROOT/bin/nvcc" >&2
+  exit 1
+fi
+
+echo "Building CUDA kernel artifacts for Linux..."
+"$CUDA_ROOT/bin/nvcc" -c "$SRC" -o "$OUT_OBJ" -I "$CUDA_ROOT/include"
+echo "Built: $OUT_OBJ"
diff --git a/build-cuda-windows.ps1 b/build-cuda-windows.ps1
new file mode 100644
index 0000000..2294d21
--- /dev/null
+++ b/build-cuda-windows.ps1
@@ -0,0 +1,20 @@
+$ErrorActionPreference = 'Stop'
+
+$msvcCl = 'C:\Program Files (x86)\Microsoft Visual Studio\2019\BuildTools\VC\Tools\MSVC\14.29.30133\bin\Hostx64\x64'
+if (-not (Test-Path (Join-Path $msvcCl 'cl.exe'))) {
+  throw "cl.exe not found at $msvcCl"
+}
+
+$cudaBin = 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.2\bin'
+if (-not (Test-Path (Join-Path $cudaBin 'nvcc.exe'))) {
+  throw "nvcc.exe not found at $cudaBin"
+}
+
+$env:PATH = "$msvcCl;$cudaBin;" + $env:PATH
+
+Write-Host "Building CUDA kernel artifacts for Windows..." -ForegroundColor Cyan
+powershell -ExecutionPolicy Bypass -File tools\build-gpudemod-kernel.ps1
+if ($LASTEXITCODE -ne 0) { throw "kernel build failed" }
+
+Write-Host "Done. Kernel artifacts prepared." -ForegroundColor Green
+Write-Host "Note: final full-app linking may still require an MSVC-compatible CGO/link strategy, not the current MinGW flow." -ForegroundColor Yellow
diff --git a/build-sdrplay.ps1 b/build-sdrplay.ps1
index d4f41e0..c652eea 100644
--- a/build-sdrplay.ps1
+++ b/build-sdrplay.ps1
@@ -3,7 +3,11 @@ $gcc = 'C:\msys64\mingw64\bin'
 if (-not (Test-Path (Join-Path $gcc 'gcc.exe'))) {
   throw "gcc not found at $gcc"
 }
-$env:PATH = "$gcc;" + $env:PATH
+$msvcCl = 'C:\Program Files (x86)\Microsoft Visual Studio\2019\BuildTools\VC\Tools\MSVC\14.29.30133\bin\Hostx64\x64'
+if (-not (Test-Path (Join-Path $msvcCl 'cl.exe'))) {
+  throw "cl.exe not found at $msvcCl"
+}
+$env:PATH = "$gcc;$msvcCl;" + $env:PATH
 $env:CGO_ENABLED = '1'
 
 # SDRplay
@@ -38,13 +42,16 @@ if (Test-Path $cudaMingw) {
 }
 
 Write-Host "Building with SDRplay + cuFFT support..." -ForegroundColor Cyan
+Write-Host "WARNING: this path still performs final Go linking through MinGW GCC." -ForegroundColor Yellow
+Write-Host "If CUDA kernel artifacts are MSVC-built, final link may fail due to mixed toolchains." -ForegroundColor Yellow
 
 $gccHost = Join-Path $gcc 'g++.exe'
 if (!(Test-Path $gccHost)) {
   throw "g++.exe not found at $gccHost"
 }
 
-powershell -ExecutionPolicy Bypass -File tools\build-gpudemod-kernel.ps1 -HostCompiler $gccHost
+# Kernel build currently relies on nvcc + MSVC host compiler availability.
+powershell -ExecutionPolicy Bypass -File tools\build-gpudemod-kernel.ps1
 if ($LASTEXITCODE -ne 0) { throw "kernel build failed" }
 
 go build -tags "sdrplay,cufft" ./cmd/sdrd
diff --git a/docs/build-cuda.md b/docs/build-cuda.md
new file mode 100644
index 0000000..454f4ff
--- /dev/null
+++ b/docs/build-cuda.md
@@ -0,0 +1,46 @@
+# CUDA Build Strategy
+
+## Problem statement
+
+The repository currently mixes two Windows toolchain worlds:
+
+- Go/CGO final link often goes through MinGW GCC/LD
+- CUDA kernel compilation via `nvcc` on Windows prefers MSVC (`cl.exe`)
+
+This works for isolated package tests, but full application builds can fail when an MSVC-built CUDA library is linked by MinGW, producing unresolved symbols such as:
+
+- `__GSHandlerCheck`
+- `__security_cookie`
+- `_Init_thread_epoch`
+
+## Recommended split
+
+### Windows
+
+Use an explicitly Windows-oriented build path:
+
+1. Prepare CUDA kernel artifacts with `nvcc`
+2. Keep the resulting CUDA linkage path clearly separated from MinGW-based fallback builds
+3. Do not assume that a MinGW-linked Go binary can always consume MSVC-built CUDA archives
+
+### Linux
+
+Prefer a GCC/NVCC-oriented build path:
+
+1. Build CUDA kernels with `nvcc` + GCC
+2. Link through the normal Linux CGO flow
+3. Avoid Windows-specific import-lib and MSVC runtime assumptions entirely
+
+## Repository design guidance
+
+- Keep `internal/demod/gpudemod/` platform-neutral at the Go API level
+- Keep CUDA kernels in `kernels.cu`
+- Use OS-specific build scripts for orchestration
+- Avoid embedding Windows-only build assumptions into shared Go code when possible
+
+## Current practical status
+
+- `go test ./...` passes
+- `go test -tags cufft ./internal/demod/gpudemod` passes with NVCC/MSVC setup
+- `build-sdrplay.ps1` has progressed past the original invalid `#cgo LDFLAGS` issue
+- Remaining Windows blocker is now a toolchain mismatch between MSVC-built CUDA artifacts and MinGW final linking
diff --git a/internal/demod/gpudemod/README.md b/internal/demod/gpudemod/README.md
index c69ee9c..adcd2a6 100644
--- a/internal/demod/gpudemod/README.md
+++ b/internal/demod/gpudemod/README.md
@@ -8,6 +8,7 @@ Phase 1 CUDA demod scaffolding.
 - `cufft` builds allocate GPU buffers and cross the CGO/CUDA launch boundary.
 - If CUDA launch wrappers are not backed by compiled kernels yet, the code falls back to CPU DSP.
 - The shifted IQ path is already wired so a successful GPU freq-shift result can be copied back and reused immediately.
+- Build orchestration should now be considered OS-specific; see `docs/build-cuda.md`.
 
 ## First real kernel
 
@@ -22,7 +23,7 @@ On a CUDA-capable dev machine with toolchain installed:
 
 1. Compile `kernels.cu` into an object file and archive it into a linkable library
    - helper script: `tools/build-gpudemod-kernel.ps1`
-2. For MinGW/CGO builds, prefer building the archive with MinGW host compiler + `ar.exe`
+2. On Jan's Windows machine, the working kernel-build path currently relies on `nvcc` + MSVC `cl.exe` in PATH
 3. Link `gpudemod_kernels.lib` 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`