Work

A desktop AI agent for local files: chat, a task board, reusable workflows, and meetings that turn into action items.

Since Aug 2026Private

Step-level learning signals for RL on Android GUI agents, built on upstream veRL.

Private

vLLM provider for Google’s LangExtract, listed in its plugin registry.

Since Sep 202530 stars

Integrated Step-Audio2, plus fixes across duplex sessions, multi-stage pipelines and CI. Now generalizing eager MTP for TTS models.

Since Dec 20257 merged PRs, 4 in review

The executor, worker and model runner of a vLLM-based engine for looped transformers.

Since 20262 PRs in review

Router models and inference runtime, and developer experience.

Since Aug 20264 merged PRs

The Triton SnakeBeta kernel for bf16 and fp16, and FP8 and INT8 quantization paths.

Since 20262 merged PRs, 2 in review