curriculum vitae · aka Lyra

Kai Cheng 程锴

AI-native builder of generative and agentic systems

1170158548@qq.comGoogle ScholarGitHub

Writing

// education

Jul 2025 – Jun 2028

Southeast University (SEU)

M.Eng. in Artificial Intelligence (professional master's) · Suzhou Joint Graduate School · PALM Lab

  • Member of PALM Lab, working on agents, vision-language models and generative AI.
Sep 2021 – Jun 2025

Tiangong University (TGU)

B.Eng. in Computer Science and Technology · School of Computer Science and Technology

  • GPA 3.82/4.0 (top 5%); President's Scholarship and Sangma Scholarship; CET-6 496.

// experience

Oct 2026 – present

Multimodal LLM Intern · Kling Team, Kuaishou Technology

Foundation Model Center · Foundation Model & Agent Algorithms Department

  • Evaluation pipelines and agentic workflows for video generation models.
Jul 2023 – Sep 2023

Full-stack engineering intern · Suzhou Tongyuan Software & Control Technology

  • Built core web front-end modules for Syslab, a next-generation scientific computing environment, using list virtualization to render large datasets.
  • Maintained a Julia socket communication library and built a VS Code extension with language-service integration.

// publications

2026

BlendEmo: Structured Set Prediction and Pair-Conditioned Ratio Modeling for Blended Emotion Recognition

Kai Cheng, Xiaohaoyang Lei, Ruiqi Wang, Haojie Li, Cunjian Chen

IEEE International Conference on Automatic Face and Gesture Recognition (FG), 2026

IEEE XploreDOIcode

Currently also working on agentic video diffusion workflows.

// projects

Aug 2026 – present

LyHarness · modular coding-agent runtime

  • Runtime: configuration, runtime state and application services are separate layers, with one path for model calls, tool dispatch and lifecycle. File and shell tools, Git worktree isolation, skills, MCP tool discovery, permission hooks, retries and background scheduling.
  • Collaboration and memory: persistent tasks and a file-based message bus for claiming work and syncing state. Persistent memory, context compaction, history recovery and spilling tool output to disk keep context small and let a run resume on another instance.
  • Workflow recovery: structured, parallel sub-tasks with stage logs, run snapshots and a persistent journal to resume after interruption; an independent evaluator checks the goal and keeps the agent going. pytest covers the task lifecycle, tool execution, memory isolation and dedup, workflow runs and goal commands.
Sep 2026 – present

Patchloom · multi-agent, self-evolving PR review and auto-fix

  • Runtime harness: task state, budgets, retries and cancellation; checkpoints to resume, node-level records for replay and failure attribution; context compressed by risky hunks, observations and token budget; four-layer memory with tenant and repository isolation.
  • Self-evolving orchestrator: a main agent dispatches security, reliability and critic sub-agents. False positives, misses, bad fixes and execution errors become prompt and skill candidates, which go live only after three gates, with dedup, versioning and rollback.
  • Evaluated on 100 controlled PR diffs: review accuracy 92.5%, risk-detection F1 from 81.9% to 92.5%, high-risk recall from 84.2% to 94.7%.
Sep 2026 – present

TencentDB-Agent-Memory · dynamic skill-queue injection (PR #1281)

  • Three injection strategies for MemoryProxy (every_queue, latest_only, adaptive_queue) using BM25 over the latest user queue, with history snapshots, incremental injection and multi-turn state recovery.
  • Compatible with the Codex and WorkBuddy protocols; immutable snapshots, a hook cache and session/queue isolation keep instances consistent. In a 100-task evaluation with 26 tests, latest_only reaches 91% success and adaptive_queue a 94.03% KV-cache hit rate.

// open source

Jun 2026 – present

TencentDB-Agent-Memory (Tencent Cloud) · community contributor

  • Public PRs on dynamic skill-queue injection, SQLite FTS5 query sanitization, checkpoint recomputation, context-cache boundaries and gateway timestamp normalization.
Aug 2025 – Oct 2025

Open Source Promotion Plan (OSPP) 2025 · OpenTiny TinyVue

Project developer · completed

  • Built TinyVue's global animation configuration and received the OSPP 2025 completion certificate.
Sep 2025 – Oct 2025

Tencent Rhino-Bird Open Source Talent Program 2025

Project developer · Outstanding Student

  • Built a TDesign Mobile Vue page template with mobile modules, internationalization and theming.

// 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

// skills

Agents & LLMs
ReAct, RAG, context engineering, MCP, agent skills, multi-agent collaboration, memory and retrieval, structured output, agent evaluation; LangChain multi-agent systems, prompt engineering; familiar with agentic RL, RLHF, DPO and GRPO
Models
PyTorch, Transformers, Hugging Face; training, LoRA and parameter-efficient fine-tuning, data processing, checkpoints and safetensors
Engineering
Python, TypeScript/JavaScript; Linux, shell, React, Vue 3, Next.js, REST APIs, CLIs, VS Code extensions, Git