Tianer Zhou
Graphics & UI Framework Engineer • OpenHarmony, Huawei
I work on the open-source graphics and UI foundations of HarmonyOS at Huawei. My focus is building high-performance, reliable rendering systems that enrich millions of consumers.
Education
Selected Coursework: Machine Learning, Distributed Systems, Operating Systems, Computer Architecture, Linear Algebra, Database, Computer Security, Web Systems.
Projects: developed basic OS kernel (threads, virtual memory, file system), implemented Paxos and sharded workers, and created a PID-controlled Arduino drone that placed in the top 5% of the course.
AI Systems & Research Projects
- JEPA World Model Research — Reproduced hierarchical JEPA from HWM_PLDM and optimized its planning/actor loop for JEPA-style world models, achieving up to 10× planning speedup while continuing accuracy-improvement experiments.
- Latent Ensemble World Models — Trained 3-member leWM ensembles for model-based planning, and trained a separate subgoal planner to guide JEPA, increasing long-horizon task success rate from 5% to 72% in PushT.
- myGPT — GPT training/inference stack built from scratch: BPE tokenization, Transformer blocks, RoPE, FlashAttention kernel, distributed KV cache, and gRPO post-training.
- deepRL — SAC, IQL, AWAC, PPO implemented from scratch. IQL reached 230/300 on Ant Maze in 30k steps.
- AI Runtime Contributions — Merged PRs to NVIDIA Dynamo and vLLM (KV-router scheduler tests, out-of-bounds fix); submitted Ironclaw setup/safety fixes and issues.
- HarmonyOS MCP Server — Developed an MCP server enabling LLM agents to inspect and control HarmonyOS device UI.
- snake-compiler — Rust x86 compiler for a JavaScript-like language with closures and dynamic typing.
Experience
- Rendering Pipeline & Shader Optimization: Developed production 2D visual-effect shaders and an asynchronous color-extraction algorithm for adaptive glass effects; built automated parameter-search tooling to tune shader parameters and improve runtime performance up to 10%.
- Model Fine-tuning and Deployment: Quantized and deployed a 6B-parameter diffusion model for mobile, generating 2K-resolution images on flagship mobile NPUs in 10 seconds. Curated a style-transfer dataset of 100+ entries and fine-tuned the model with LoRA on 4 V100s using FSDP.
Exploring on-device 3D Gaussian rendering and lightweight generative models for mobile.
Core contributor (1,200+ commits, ranked Top 1 among 1,300+ contributors) to the open-source UI engine. Built high-performance lazy/reusable scroll containers that reduced jank rate by 60%, designed deduplicated image caching, and created a React Native SVG renderer 20% faster than the Android reference.
Technical Skills
- AI Systems: Diffusion / flow matching, LLM, JEPA, Model-based Planning, RL, training / inference parallelism, KV Cache, attention kernels
- Frameworks: PyTorch, Jax, Lightning
- Programming Languages: C++, TypeScript, Rust, Go, Python
- Systems: Distributed Systems, Graphics Runtime, Compiler, xv6 Kernel
- AI Harness: Codex, Claude Code, Openclaw