High-performance C++ pipeline & heterogeneous inference framework for Edge AI & LLMs, featuring fixed-batch DMA scheduling, multi-modal DAG orchestration, and pure C ABI isolation.

Open-source CPU inference and PyTorch tools for Needle 2: CQ2/CQ4 quantization, model conversion, QAT, and ARM NEON/SDOT kernels.

A multilingual-to-Chinese translation evaluation platform with LLM-based scoring, model comparison, and evaluation tracking.

Extensible output-vocabulary pruning for decoder-only LLMs while preserving multilingual input embeddings. Tested on Qwen2/Qwen2.5 and Gemma 3.

A ~100M parameter Qwen3-based multilingual-to-Chinese machine translation model, trained from scratch with an automated AutoResearch pipeline.