Android 16 fork. AI as a platform primitive. Twelve capabilities, one shared runtime, every app. OEM-pluggable. Apache 2.0.

Inference-aware runtime for AI coding agents that reuses execution state to reduce repeated reasoning, repo rereads, tool calls, and failure loops.

Open source Node.js runtime for local LLM inference, on-device AI, and private model execution.

High-performance Multi-LoRA inference runtime for serving multiple fine-tuned adapters over a shared base model. Features sub-3ms hot-swapping latency and ~2MB memory footprint per adapter. Optimized for multi-tenant LLM serving.

Multi-provider LLM runtime core: routing, key management, and resilient fallback execution for agent orchestration.

JibarOS organization profile.

Single-GPU LLM decode research prototype: paged KV cache, Triton attention, CUDA append, scheduling, shared prefixes, and multi-layer transactions.

Shared JavaScript runtime core for local LLM inference across Node.js, browser, and React Native.

从零实现的 C++/CUDA 张量库与 LLM 推理系统学习项目，包含 CUDA Kernel、KV Cache、PagedAttention、Prefill/Decode、Continuous Batching 与 Nsight Profiling。

High-performance, local-first RAG (Retrieval-Augmented Generation) stack. Features PDF/TXT ingestion, ChromaDB vector storage, and Qwen2-based generation via FastAPI. Optimized for sub-2s latency on consumer GPUs (8GB VRAM).

A domain-agnostic, production-oriented, high-performance Rust video perception runtime built on GStreamer.

Portfolio: platform security and AI infrastructure engineering