Projects

Built an LLM planner that turns natural-language instructions into consecutive and executable operations for an ABB GoFa robot.

• Decomposes instructions into multi-step task lists and coordinates parallel robot operations with execution-status tracking.
• Built an MCP server over Robot Web Services and RAPID for equipment queries, motion, picking, placing, and tool-call feedback.
• Added execution-failure handling and short-term context so the planner can recover and continue the workflow.

Created a vision-language-action agent for voice-guided UR robot grasping and tool-delivery tasks for the 12th NVIDIA Sky Hackathon.

• Converted spoken requests into real-time task plans, tool selections, and robot actions through a LangGraph agent.
• Combined FastSAM and vision-language models for scene understanding with Strategy of Mind prompting and GraspNet for grasp planning.
• Exposed UR robot actions through MCP tools for collision-free pickup and delivery.
• Our project received the Excellence Award (National Second Place).

Built a competition robot that autonomously transported and placed objects along designated routes.

• Designed the mechanical structure and integrated sensing, actuation, and control.
• Coordinated DC motors, encoder motors, servos, and robotic arms through an STM32 controller and CAN bus.
• Added a dynamic shortest-path algorithm and moved from one to three motors, cutting operation time from 4 min 32 s to 3 min 15 s.
• Our project received the National First Place Award.

Built a controlled agent harness that turns visual references, wireframes, and content into responsive web pages with isolated browser previews.

• Orchestrates visual analysis, design, HTML implementation, validation, review, and preview publication as one observable workflow.
• Delegates bounded work to design, implementation, and quality-review subagents, streaming plans, reasoning, tool calls, and approvals.
• Relies on thread-scoped state, Skills, filesystem isolation, retries, call budgets, and deterministic HTML quality gates for verifiable artifacts.

Developed a configurable workstation that binds language models, manufacturing knowledge, digital assets, and production-line operations into reusable agent workflows.

• Configures agents with roles, goals, models, knowledge bases, digital assets, and assigned tasks.
• Composes agents and Crews in a visual workflow with retrieval, model-acquisition, parsing, and equipment-control tools.
• Connects to production-line devices over WebSocket for control, database queries, data analysis, and monitoring.

Created a browser-based town where configurable AI characters keep context and interact with users and one another.

• Lets users define NPC names, demographics, and personalities, and manage relationships between characters.
• Integrates LangChain agents, contextual memory, and retrieval-augmented reasoning into a Phaser environment.
• Supports multiple maps with import/export of reusable world and character configurations.

Built an AI review workspace that turns real Bilibili danmaku and comments into evidence-backed insights for creators.

• Accepts a video URL, BV ID, or b23.tv short link, retrieves the real feedback, and shows the data distribution before analysis.
• Streams the analysis to extract high-frequency topics, sentiment, valuable feedback, and actionable content suggestions.
• Links every insight back to the original comments or danmaku, with views for user levels, danmaku density, sentiment timelines, and priority feedback.

Built a research website for organizing literature, exchanging technical ideas, and discussing questions with specialized LLM agents.

• Lets researchers curate their top 100 papers and organize discussions by technical board, with viewpoints, comments, and links back to source material.
• Includes configurable research agents and a fine-tuned digital-twin assistant for deeper academic question answering.

Built an AI-assisted messaging and work-order system that answers rental customers and helps staff manage service requests.

• Receives WhatsApp messages through a Twilio webhook and synchronizes them with the backend in real time.
• Gives staff a WebSocket workspace to review the conversation and agent reply, and turns each request into a traceable work order with status, priority, property, and approval details.

Developed an in-browser AI assistant that combines webpage context, visual understanding, search, and personal knowledge in one conversation panel.

• Answers questions against the active webpage without moving content into a separate chat application.
• Uses a vision-language model for image recognition and visual question answering, plus web search and knowledge-base access.