Text generation
Local GGUF model loading, streamed generation, persistent sessions, and model-native chat formatting through a small Python API or an OpenAI-compatible local server.
Technical foundation
Seven modalities. One local runtime.
SessionNest Runtime is a Vulkan/WGSL AI runtime built for LessPrep. It keeps model work on a compatible Windows computer and gives the apps one controlled path for text generation, images, speech recognition, speech synthesis, vision, embeddings, and subject cutouts.
All seven modalities
Image generation
Local SD 1.5 and SDXL-class generation with process isolation, cancellation, device-aware memory limits, and intermediate previews for app workflows.
Speech-to-text
Local Whisper transcription turns recorded speech into text without requiring a hosted speech service.
Text-to-speech
Local SpeechT5 synthesis produces spoken output through the same runtime surface.
Vision
Local vision-language processing supports image understanding for multimodal application workflows.
Text embeddings
Local embedding models create semantic vectors for search, retrieval, and related-text matching.
Background removal
A native U²-Net segmentation path isolates a subject and returns a transparent RGBA cutout. It runs on SNRT’s own Vulkan kernels and loads only when requested.
Why build it locally?
LessPrep handles learner materials and clinical drafts. Local processing reduces the need to send that work to a hosted model, while explicit cancellation, memory checks, and review-first product rules keep the human in charge.
Hardware support and performance vary. SNRT is still under active development and is evaluated on both discrete GPUs and lower-power integrated graphics.
Technical questions
Ask for the deeper notes on supported models, validation, memory controls, and the native background-removal implementation.
lessprepinfo@gmail.com