Know exactly what local AI your Windows machine can run — in 30 seconds.
Pure PowerShell. No installs. No dependencies. No telemetry. Your data never leaves the machine.
Ready to deploy? The scanner generates a hardware blueprint. We turn it into a working local AI stack — Ollama, Open WebUI, Flowise, and n8n — configured exactly for your machine. Get your custom deployment script — $45 → · Full managed deployment →
Most local AI guides tell you to install Ollama. None of them tell you what model your specific hardware can run without stuttering, which quantization fits in your VRAM, or how to split transformer layers across your GPU and CPU.
Cascade Scanner does the math. One command. One YAML file. A three-model stack ready to deploy:
| Role | What it's for |
|---|---|
| General | Private assistant, document analysis, multi-step reasoning |
| Code | Code generation, debugging, refactoring |
| Embedding | RAG — semantic search over your own documents |
Every model in the stack is calculated to fit your exact VRAM headroom. The YAML it produces is machine-readable by any Cascade-spec-compliant deployer — including the $45 automated deployment script.
No installation. No admin rights required. No data sent anywhere.
# Download cascade-scanner.ps1, then:
powershell -ExecutionPolicy Bypass -File .\cascade-scanner.ps130 seconds. One YAML file. Done.
# Custom output path
.\cascade-scanner.ps1 -OutputPath "C:\my-config.yaml"
# Silent mode (YAML only, no console output)
.\cascade-scanner.ps1 -SilentConsole summary:
======================================================
CASCADE SCANNER - ASSESSMENT RESULTS
======================================================
GPU : NVIDIA GeForce GTX 1650
VRAM : 4GB total / 3072MB usable
CPU : Intel(R) Core(TM) i7-10750H CPU @ 2.60GHz
Threads : 6 cores / 12 threads @ 2.59GHz
RAM : 64GB
--- DEPLOYMENT RECOMMENDATION ---
Primary model : llama3.1:8b-instruct-q4_K_M
GPU layers : 19 / 32
Inference mode : hybrid-gpu-cpu
Speed estimate : 8-15 t/s
Stack tier : full (ollama, open-webui, flowise, n8n)
--- MODEL STACK ---
[GENERAL] llama3.1:8b-instruct-q4_K_M
19 / 32 layers hybrid
[CODE] deepseek-coder:6.7b-instruct-q4_K_M
25 / 32 layers hybrid
[EMBEDDING] nomic-embed-text
12 / 12 layers full-gpu
Total disk required: 8.8 GB
--- UPGRADE PATH ---
RTX 4060 Ti 16GB (~$450 USD) - unlocks 13B full precision on GPU
YAML output (excerpt) — see examples/sample_config.yaml for the full file:
ollama:
primary_model: "llama3.1:8b-instruct-q4_K_M"
inference_mode: "hybrid-gpu-cpu"
num_gpu_layers: 19
environment:
OLLAMA_NUM_GPU: "19"
OLLAMA_NUM_THREAD: "10"
OLLAMA_MAX_LOADED_MODELS: "2"
OLLAMA_KEEP_ALIVE: "5m"
models_to_pull:
- role: "general"
model: "llama3.1:8b-instruct-q4_K_M"
gpu_layers: 19
disk_gb: 4.7
- role: "code"
model: "deepseek-coder:6.7b-instruct-q4_K_M"
gpu_layers: 25
disk_gb: 3.8
- role: "embedding"
model: "nomic-embed-text"
gpu_layers: 12
disk_gb: 0.27
total_disk_required_gb: 8.8The scanner uses pure Windows Management Instrumentation (WMI) — no external dependencies, no installs, no admin rights required in most cases.
- GPU detection — Queries
Win32_VideoController, ranks adapters by VRAM, filters virtual and Microsoft display drivers - VRAM correction — Reads the registry QWORD first (avoids the
uint32overflow in WMI that silently caps any GPU above 4 GB). Falls back to a correction table of known cards, then WMI as a last resort - CPU + RAM + storage — Standard
Win32_Processor,Win32_ComputerSystem,Win32_LogicalDiskqueries - Layer fit calculation — For each model:
floor(usable_vram_mb / mb_per_layer), capped at the model's total layer count. Usable VRAM is 75% of detected VRAM to reserve headroom for KV cache - Model stack selection — Picks a general model based on RAM and disk tiers, adds a code model when disk allows, always includes a embedding model for RAG
- YAML emission — Writes the result conforming to the Cascade Deployment Spec v1
This scanner is the reference implementation of the Cascade Deployment Specification v1 — a structured YAML format for describing local AI deployments on Windows. The full spec is in docs/SPEC.md.
The spec is designed to be consumed by a deployment script. Every field the deployer needs is present: GPU layer counts, Ollama environment variables, model identifiers with roles, stack components, port assignments, and total disk required. If you're building a deployer, dashboard, or alternative scanner, your tool can interoperate with any Cascade-spec-compliant software.
The math for running local AI is genuinely annoying: VRAM minus headroom, divided by megabytes-per-layer at your quantization, capped by total transformer layers — and Windows under-reports VRAM on consumer cards, so most tools get it wrong.
Cascade Scanner encodes that math correctly so you don't have to. We open-sourced the assessment because it shouldn't be the moat. The deployment, the workflows, and the ongoing operation are. If you want to run the assessment yourself and self-deploy, this tool is yours, free, forever.
Cascade Scanner is the free entry point to the Cascade Service by Fourier Systems — local AI for regulated small businesses that cannot put client data in the cloud: law firms, medical offices, financial advisors, accounting practices.
The YAML file the scanner produces contains everything needed to stand up a working local AI stack. Upload it at checkout and we generate a PowerShell deployment script built exactly to your hardware profile — your GPU layer count, your model, your quantization, your ports. Run it once. Your stack is live.
Not a template. Not blanks to fill in. A script that knows your machine.
Checkout link coming soon — contact Fourier Systems to get early access.
Send us your YAML and we return a branded PDF containing your hardware capabilities in plain business language, a specific upgrade recommendation with dollar figures, and a 30-minute consultation call.
Everything the scanner identifies, fully deployed and supported:
- Ollama + Open WebUI + Flowise + n8n stack
- Five pre-built business automation workflows (document intake, financial briefings, client inquiry routing, competitive research, invoice processing)
- NVIDIA NIM cloud overflow for heavy workloads
- Cascade Certification compliance attestation
- Optional NVIDIA performance tier (vLLM, RAPIDS, Triton, NIM containers)
For regulated SMBs ready to move: contact Fourier Systems.
Issues and pull requests welcome. See CONTRIBUTING.md for guidelines.
Especially useful contributions:
- GPU model entries for the VRAM correction table
- New model definitions (Ollama identifier, MB-per-layer, total layers, disk size)
- Linux/macOS port of the scanner (requires a WMI replacement)
- PDF report generator that consumes the Cascade Spec v1 YAML
MIT — see LICENSE.
Fourier Systems (17630646 Canada Inc) is a Toronto-based technology services company specializing in local AI deployments for regulated small businesses. We build the Cascade Service: hardware-optimized, privacy-first, on-premise AI for organizations that cannot put their data in the cloud.
- Website: coming soon
- Contact: coming soon
- Cascade Service: From $800 deployment + $99/month support
Run the scanner free. Self-deploy, or let us do it. Either way, your data stays on your machine.