Research / Graduation Thesis

Research on a Collaborative Multi-Agent Platform

I designed, implemented, and evaluated an autonomous system that converts ambiguous natural language instructions into executable tasks across web operations, IoT control, knowledge retrieval, and scheduling.

LangGraph MCP RAG Docker Flask / REST API Jetson / Raspberry Pi

Research Presentation

Slides (English): NCSP-Presentation-EN.pptx

Open Presentation

Research Summary in 30 Seconds

Problem Solved

Built a system that can execute ambiguous requests like "do it as usual" across both web and IoT tasks.

Architecture

Integrated a memory-enabled orchestrator with four specialist agents (Browser / IoT / Life-Style / Scheduler).

Implementation Strength

Designed a robust Plan → Execute → Review loop with automatic retries for failed tasks.

Novelty

Standardized capabilities via MCP and implemented hierarchical inference using both cloud and edge LLMs.

Quantitative Outcomes (Thesis Evaluation)

Metric	Result	Impact
Integrated score (10 scenarios)	15 (no memory) → avg 25.0 (with memory)	About 1.7x improvement (+67%)
Follow-up questions	Baseline total 2 → memory-enabled personas total 0	Lower user burden
Edge inference speed	CPU 9.64 tok/s → GPU 24.76 tok/s	About 2.57x faster
Personalization example	Used address, allergy, and preference data to automate search/recommendation/scheduling	Demonstrated practical context awareness

Career-Relevant Contributions

Requirements design: structured ambiguous user intents into executable task flows using memory-driven context.
Backend implementation: decoupled agents with REST APIs and operated them as Docker-based microservices.
Reliability engineering: implemented a self-correction loop that reviews outputs and retries failed actions.
AI engineering: combined Function Calling, RAG, and model routing to balance quality and cost.
Evaluation design: validated improvements with both component-level tests and 10 scenario-based end-to-end tests.

Tech Stack and Practical Strengths

AI / Agent Systems

LangGraph, MCP, Function Calling, RAG (LangChain + FAISS), dynamic model routing.

Backend / Infrastructure

Python, Flask, REST API, SSE, SQLite, Docker, async job queue, Chrome CDP.

Edge Integration

Integrated Jetson Orin Nano, Raspberry Pi 4, and Raspberry Pi Pico W with cloud-edge role separation.

Professional Value

End-to-end execution from ambiguous requirement analysis to API design, evaluation, and performance improvement.

Known Issues and Next Improvements

For knowledge-cutoff time drift, proposed strict timestamp injection per task to keep time references accurate.
For complex web UI failures, planned multimodal screen understanding in addition to DOM-based control.
Prioritized a continuous improvement cycle across accuracy, reproducibility, operational stability, and cost.