Core technology

Engineering first: measurable, regression-tested, and rollback-ready.

Large model stack

The training side uses mainstream distributed frameworks and parameter-efficient fine-tuning (LoRA / adapters). Inference is split by workload: batch documents use offline queues; online chat uses rate limits and caching. Outputs are validated with a unified JSON Schema, with automatic retries and fallback templates on failure.

Multimodal fusion

Text, vector primitives, raster images, and tables are weakly aligned in one project space; cross-page references use frame IDs and layout trees to avoid plausible but wrong cross-document bleed.

Edge computing architecture

Vision models export to ONNX / TensorRT. Edge and center sync metrics and sample fingerprints only; raw images stay on premise by default.

Data security and compliance

Private knowledge bases can run fully offline. When public LLM calls are required, we support field-level desensitization and a configurable "zero egress" mode with a local small model as fallback.

View architecture overview Read white paper summary