This is the shortest practical path to boot the current project on a local Windows machine.
From the repo root:
.\setup.ps1
ollama pull qwen2.5:3b-instruct
ollama serve
.\start.ps1This is the default local Ollama path. If you are using a hosted or remote Ollama-compatible endpoint instead, keep .\setup.ps1 and .\start.ps1, but set OLLAMA_URL, OLLAMA_MODEL, optional backend-only OLLAMA_AUTH_TOKEN, and OLLAMA_MODE=remote before startup. That remote path changes normalization only; ASL inference still runs on this machine.
If you want to skip Ollama normalization entirely and work in raw gloss mode:
.\setup.ps1
.\start.ps1 -NoOllamaThen open:
http://127.0.0.1:5500/index.html
If landing/dist/index.html exists, http://127.0.0.1:5500/ will open the landing site and the app will remain available at /index.html.
In the UI:
- Choose
ASLfor the simplest first run. - Accept the in-app consent prompt.
- Press
Start.
ASL mode runs inference in the backend started by start.py. The current frontend default sends browser camera frames to that backend, and backend-local camera mode still exists as a fallback. Voice mode is the path that needs browser microphone permission.
If you want Voice mode, finish the credential and OAuth setup in INSTALLATION_GUIDE.md first.
setup.ps1:
- finds Python on
PATH - creates
.venvif needed - auto-detects
cudavscpuwithnvidia-smi - installs the matching runtime requirements
Examples:
.\setup.ps1
.\setup.ps1 -Profile cuda
.\setup.ps1 -Profile cpuProfile meanings:
auto: choose CUDA when available, otherwise CPUcuda: force a CUDA attemptcpu: force CPU-only install
start.ps1:
- sets
ASL_RUNTIME_PROFILE - activates
.venv - runs
start.py - starts the frontend plus ASL, Voice, Letter, and Admin services
Examples:
.\start.ps1
.\start.ps1 -Profile auto
.\start.ps1 -Profile cuda
.\start.ps1 -Profile cpu
.\start.ps1 -NoOllama
.\start.ps1 -TunnelDefault service URLs:
- Frontend root:
http://127.0.0.1:5500/ - Main app:
http://127.0.0.1:5500/index.html - ASL API:
http://127.0.0.1:8000 - Voice/Auth API:
http://127.0.0.1:8001 - Letter API:
http://127.0.0.1:8002 - Admin API:
http://127.0.0.1:8003
Once the stack is running, check the ASL hardware profile:
Invoke-RestMethod http://127.0.0.1:8000/runtime | Format-ListLook for fields such as:
profilesourcetorch_cudaonnx_cudaonnx_providersgpu_namedirectml_supported
Expected behavior:
- NVIDIA machine with full CUDA support:
profileshould becuda - Non-NVIDIA or incomplete CUDA install:
profileshould becpu directml_supportedis currentlyFalse
If you used the default CUDA setup:
.\.venv\Scripts\python.exe -m pip install -r requirements.test.txt
.\.venv\Scripts\python.exe -m pytestIf you installed the CPU profile:
.\.venv\Scripts\python.exe -m pip install -r requirements.test-cpu.txt
.\.venv\Scripts\python.exe -m pytestCheck these local prerequisites first:
- checkpoint file available at the default path, or
ASL_CHECKPOINT_PATHset - reachable Ollama endpoint available for normalization (
OLLAMA_URLlocal by default, or a hosted endpoint if configured) - webcam connected for
ASL - microphone connected for
Voice - Google and Azure credentials configured if you want
Voicemode