Our team comes from the University of Toronto Machine Learning Group and the Vector Institute. We work on AI safety, agentic coding, and Chinese-language NLP — and turn the results into deployable, auditable systems.
Our August frontend update introduces Yanlan-expand: single-character differences become phrase-level highlights, keeping the context of each correction in view.
Read full
All 8 weighted metrics ahead of V2.0, with 43 of 47 comparisons won. Chinese pre-publication correction for publishing, government communications, and enterprise content.
Read full
Jointly optimising reasoning and self-refinement under one binary reward — no critic model, no process reward, no critique data. Adds +11.5 pp on AIME pass@4 (Qwen3-4B) over vanilla GRPO.
Developed with Xiaohongshu, this mobile coding-agent benchmark includes 50 production iOS feature tasks, PRDs, Figma specifications, and 449 hand-written tests; the best-performing evaluated agent-model pair completes 12%.
A folder-scoped coding agent, deployed entirely inside your own network. Coolwei Code reads and edits files, runs commands under a boundary you set, and puts every change up for review — in a macOS application and at the command line, against the model gateway you already operate.
A multi-stage RL system for pre-publication correction across text, subtitles, transcripts, and OCR, tuned for false-alarm control and deployment efficiency.
A runtime safety system for AI services. RedShield scores risk while the answer is being written and cuts the output off when it crosses the line you set; TingLan watches word-level cues inside the model and cuts the output off before the risky content is ever spoken — the cues then go to human review. The main model's weights are unchanged.
Hand it a deck and a script and it presents page by page, answers questions automatically, and takes interruptions mid-flow. Built for service halls, showrooms, briefings and corporate training, with on-premises deployment.
Coolwei AI Lab is the research arm of Coolwei (网景盛世), based in Jinzhou, with a team from the University of Toronto machine learning group and the Vector Institute. We publish on AI safety, coding agents, reasoning, evaluation and Chinese-language NLP at venues including KDD, EMNLP, COLM and NeurIPS, then turn that work into systems that run inside publishing, government and enterprise workflows.
