About me

I'm Gan Liu (刘淦) — a founding engineer and 3D-vision researcher based in Shenzhen. I like working at the boundary between hard systems problems and products people actually use, and I try to own the whole path between them.

My research is in real-time 3D reconstruction. With collaborators at SUSTech and the Jiang Liu lab, I worked on depth estimation and SLAM for ophthalmic and endoscopic surgery — settings where there is no depth sensor, the tissue deforms, and classical geometry drifts. That line of work led to a paper at CVPR 2026 (RetinalDepth-64K) and the SurgFM-SLAM project on adapting 3D foundation models.

As a founding engineer at Citronetic, I work on post-training data and environments for AI agents. We build simulated workplaces where a model has to read the inbox, dig through chats and docs, and write back to the right person — and where every task can be graded, so the same environment that evaluates an agent can also train it. I've also built the agent infrastructure underneath: unified MCP + REST contracts, multi-agent orchestration through a shared board state machine, and model routing with graceful degradation.

I care about clarity and craftsmanship: the contract that stays stable, the fallback that never hard-fails, and the small details that make software feel fast and alive.