Lee Ho

LeeAgent: A Local Coding Agent for Long Tasks and Multi-Agent Collaboration

• Designed a client-server architecture with persistent TCP connections for command calls and event delivery between the CLI/TUI and a backend service. The backend executes model calls and tool scheduling asynchronously, independently of client connections.
• Built a typed event bus that streams model output and tool events by task and session, with history replay, live subscriptions, and event recovery and deduplication after reconnects.
• Wrapped an OpenAI-compatible model layer and implemented a ReAct-style agent loop with unified tool scheduling. Integrated file read/write, Bash execution, and task-list tools, supporting multi-round tool use and self-correction from error feedback.
• Added session persistence, history resumption, and context compression for long tasks. Built parent-child agent collaboration with separate contexts and event buses, event bridging, result reporting, Markdown Skills loading, and MCP tool integration.
• Established three trace layers across IPC, domain events, and model calls, tracking each run and turn together with token usage and elapsed time.

Connect6 Game System with an Attention-Enhanced Policy-Value ResNet

Led the technical design, model architecture, and experimental validation of a Connect6 decision-making system.

• Encoded board states with multiple channels. Residual modules with 32, 64, and 128 channels extracted spatial features, while spatial and coordinate attention modeled critical regions.
• Built policy and value heads to predict the next-move distribution and the likely outcome of the current board position.
• Achieved 73.42% policy top-1, 91.49% policy top-5, and 0.221 value MSE on the validation set. Ranked first with an 86.2% win rate in Jilin University's Connect6 ranking tournament and won National First Prize in the 2025 Chinese Collegiate Computer Game Competition.

LLM Reasoning Optimization with Reinforcement Learning and Chain-of-Thought

Proposed a GRPO-based reinforcement learning framework with multi-granularity rewards to address unsupported emotion inferences by VideoAuto-R1 when audio is missing at test time.

• Designed an explicitly segmented episode-level reasoning chain to structure the model's reasoning process.
• Designed semantic calibration rewards that distinguish unsupported fabrications, reasonable hypotheses, and visually grounded inferences, improving semantic consistency when audio is unavailable.
• Reduced reliance on the final answer during training by rewarding sound reasoning even when the outcome is wrong. Added stability rewards based on hidden-state perturbations to improve generalization and reliable reasoning.