Platform Features
Engineered for developers who demand enterprise data sovereignty with the convenience of a modern web interface.
Algorithmic Auto-Tuner
No more guessing hyperparameters. Auto Mode calculates optimal rank, alpha, learning rate, micro-batch size, and sequence length with 15-20% guaranteed VRAM headroom.
QLoRA & 4-bit Quantization
Train 7B-14B models on consumer GPUs (12GB-16GB VRAM) using 4-bit NF4 quantization, gradient checkpointing, and paged AdamW optimizers.
Corpus Quality Scoring
Transparent 0-100 heuristic scoring engine that detects duplicate records, near-duplicates, encoding noise, and anomalous sample lengths before you waste compute.
Local RAG Studio
Chunking presets (code, structured documentation, narrative) and dual-stage retrieval (candidate top-k and rerank k) stored entirely in a local vector database.
Multi-Backend Hardware Engine
First-class support for NVIDIA CUDA on Linux and Windows, as well as Apple Silicon unified memory via the native MLX framework.
Zero-Leak Privacy Protocol
Local filesystem paths, raw documents, and adapter weights are kept strictly on-device. The server only sees opaque identifiers and loss telemetry.