I'm building KineWorld, a research-stage world-model company.
My current focus is action-conditioned prediction, compact latent dynamics, and reproducible evaluation for physical intelligence. I also build the agent tooling and security infrastructure that make experiments easier to inspect and repeat.
勘境 / KineJing is our world-model integration and research project. Start with the code, then follow the model cards and experiment records.
| Project | What to explore |
|---|---|
| KineJing · 勘境 | CPU motion baseline, model adapters, and a trained three-view action-conditioned feature predictor |
| Kine-JEPA | Compact latent-model prototypes, action-conditioned rollout, and planning interfaces |
| KINE-Bench | Evaluation protocols, representation diagnostics, baselines, and recorded negative findings |
| KINE-DataPipe | Video preprocessing, motion filtering, event-candidate mining, and pair construction |
- Try the CPU workflow: KineJing quick start
- Inspect the trained predictor: model card
- Check the experiments: evidence records
- Understand the organization: KineWorld · website
Current results are internal and research-stage. The trained KineJing predictor outputs future visual features; it has no RGB decoder. Its model card records the data split, baselines, weight hashes, and limitations.
The trained predictor checkpoint is currently retained in a private company release. The public repository provides the CPU demo, training code, and evaluation records; reproducing the trained predictor requires checkpoint access and separately obtained upstream data and weights.
Public demos, software tests, and model-quality evaluations answer different questions. I keep those distinctions visible and welcome independent reproduction, including failed attempts.
My agent and security work includes KineGrant Protocol, Developer Intelligence, Agent Plugin Doctor, and Agent Replay.
These projects cover capability-based authorization, developer tools, plugin validation, and agent observability.
For world-model research, useful starting points are a reproducible failure case, a baseline comparison, or a clearly scoped experiment. See KineWorld's contribution guide for evidence requirements and how to contribute.