IEEE FG 2026
~/lyra-planet · Kai Cheng 程锴
Lyra
I build generative models
- Multimodal LLM InternKling Team · Kuaishou Technology
- M.Eng. in AISoutheast University · PALM Lab
- FG 2026first-author paper
lyra@planet ~ % claude usage --daily
live from my session logs · how I count
View as table
// at a glance
// journey
Tiangong University
B.Eng. in Computer Science · GPA 3.82, top 5%
Tongyuan
Full-stack intern · Syslab front end, VS Code extension
Southeast University
M.Eng. in AI · PALM Lab, Suzhou
Open source
OSPP · OpenTiny TinyVue; Tencent Rhino-Bird outstanding student
BlendEmo at FG 2026
First-author paper, IEEE FG
Kling Team, Kuaishou
Multimodal LLM intern · Foundation Model Center
// selected work
GalForge
lyra-planet/galforge
An Agent Skill that turns one sentence into a playable pixel-art visual novel. Sprites, scenes and music are all written as code, and every game must pass ten bundled checks, including an exhaustive ending solver.
newPlay in the browser ↗Source ↗Read the write-up →
JavaScriptagent-skillvisual-novelpixel-artwebaudio
LyHarness
lyra-planet/LyHarness
Modular coding-agent runtime: tools, worktree isolation, skills, MCP, persistent tasks and memory, resumable workflows.
Pythonagentscoding-agentmcp
Patchloom
private repo
Multi-agent PR review and auto-fix with a self-evolving orchestrator; 92.5% review accuracy on 100 controlled PR diffs.
multi-agentcode-reviewself-evolving
TencentDB-Agent-Memory
TencentCloud/TencentDB-Agent-Memory · PR #1281
Dynamic skill-queue injection for an agent memory proxy: three BM25-based strategies, 91% task success in a 100-task eval.
TypeScriptagent-memoryretrievalopen-source
BlendEmo
lyra-planet/BlendEmo
Code for our FG 2026 paper: blended emotion recognition via structured set prediction and pair-conditioned ratios.
Pythonmultimodalemotionpytorch
GAL-Dreamer
lyra-planet/GAL-Dreamer
Multi-agent system that turns one conversation into a complete galgame: world-building, characters, story outline and consistency checks.
Pythonmulti-agentlangchainllm
Pipline
lyra-planet/Pipline
Multi-stage video editing with MiniMax-H3 through a hosted API or a local ComfyUI workflow, with retries and failure repair.
Pythonvideo-gencomfyuipipeline
// publication
Currently also working on agentic video diffusion workflows.
// open source
- TencentDB-Agent-Memory (Tencent Cloud) · community contributor
- Open Source Promotion Plan (OSPP) 2025 · OpenTiny TinyVueProject developer · completed
- Tencent Rhino-Bird Open Source Talent Program 2025Project developer · Outstanding Student
// honors
Lanqiao Cup Digital Technology Innovation Contest, national finals: 2nd place nationwide (National First Prize)
Lanqiao Cup Software Contest, Web Application Development (university group): National First Prize
RAICOM Robot Developer Competition (CAIP): National First Prize
China Collegiate Computing Contest (Computer Design Competition): National Third Prize
// stack
- PyTorch
- Transformers
- Hugging Face
- LoRA
- diffusion
- ComfyUI
- vision-language models
- MCP
- agent skills
- LangChain
- ReAct
- RAG
- context engineering
- agentic RL
- GRPO
- DPO
- Claude Code
- Python
- TypeScript
- React
- Vue 3
- Next.js
// writing
- GalForge: a visual novel from one sentence
GalForge is an Agent Skill. Ask your coding agent for a game in one line, and it writes, draws, scores and tests a branching visual novel that runs in the browser. This is the game it made from a one-line request, and how it checks its own work.
- A planet that denoises itself, in pure CSS
The banner on my profile loops from noise to a pixel planet and back. There is no JavaScript: two noise layers, one opacity schedule, and a few rules about pixel grids that I learned the hard way.
- How I count my Claude Code tokens
Summing every usage field in the session logs gives a number that is wrong in both directions. Here is what I count instead, and how close it lands to ccusage.