Shigeki Noguchi
Keio University, B.A. Economics
I'm a passionate researcher and developer specializing in cutting-edge AI systems and full-stack software engineering. My work focuses on pushing the boundaries of what's possible through multi-agent systems, LLMs, and innovative web applications.
ML/AI
Multi-Agent Systems, LLMs, Evolutionary Algorithms
Full-Stack Dev
Django, Python, HTML, CSS, JavaScript, Node.js, Swift
AI Research
Research Experience in a Multinational Team
Projects
Hierarchical Multi-Agent System
Revolutionary AI Performance Enhancement
Whitepaper is available at Here
Developed a groundbreaking hierarchical multi-agent debate system that achieved remarkable performance improvements in AI reasoning tasks. The system elevates smaller language models to match the performance of much larger models through sophisticated agent collaboration.
Key Achievements:
- 🎯 Performance Breakthrough: Achieved 90-92.5% accuracy vs 70% baseline
- 🧠 Smart Architecture: Dynamic hierarchy parsing system
- ⚡ Efficiency Innovation: 20%+ improvement in reasoning accuracy
- 📊 Proven Results: Validated on mathematical reasoning benchmarks
Performance Results (GSM8K Dataset):
Multi-Agent System
92.5-90.0%
Zero-Shot Accuracy
Single Agent
71.5-69.8%
Zero-Shot Accuracy
*OpenAI GPT-4o-mini is used for the evaluation.
Technologies: Microsoft AutoGen, Python, Multi-Agent Architectures, Collective Intelligence, GPT-4o-mini
Impact: Demonstrated the possibility of achieving AGI-level performance with Multi-Agent Systems where smaller language models can outperform larger models.
Multi-Vendor E-Commerce Platform
Modern and Secure Web Application
Built a comprehensive multi-vendor e-commerce platform supporting multiple sellers, secure payment processing, and advanced marketplace functionality.
Key Features:
- 🛍️ Multi-Vendor Support: Comprehensive seller management system
- 💳 Secure Payments: Integrated payment gateway with Stripe
- 🔐 Authentication: Advanced user management and security
- 📱 Responsive Design: Responsive design for all devices and Native App by Swift for iOS
- ✨ AI Adaptation: AI-powered product management system
Technologies: Django, Python, Stripe API, HTML/CSS/JavaScript
Impact: Production-ready e-commerce solution demonstrating full-stack development capabilities and understanding of complex business requirements.
Real-Time Collaborative Whiteboard
Available at postit.space
Created a sophisticated real-time collaborative whiteboard application with advanced drawing capabilities and multi-user synchronization. The application supports instant collaboration across multiple users with professional drawing tools.
Technical Highlights:
- ⚡ Real-Time Sync: Instant collaboration across multiple users
- 🎨 Advanced Drawing: Vector-based drawing with professional tools
- 🔄 State Management: Efficient real-time data synchronization with Liveblocks
- 📱 Cross-Platform: Responsive design for all devices
Technologies: Next.js, TypeScript, Tailwind CSS, Liveblocks, Tldraw, Cloudflare, Vercel
Impact: Showcases modern web development skills with focus on user experience and real-time collaboration technologies.
Marvin Minsky's "Society of Mind" inspired Multi-Agent System
Mathematical Problem-Solving with Specialized Agents - Available on GitHub
Inspired by Marvin Minsky's "Society of Mind," developed a multi-agent system that decomposes mathematical problem-solving into specialized agents: Interpreter, Planner, Solver, Verifier, and Answer Generator. Each agent handles a specific aspect of the reasoning process.
Specialized Agents:
- 🔍 Interpreter: Extracts key information from problem statements
- 📋 Planner: Devises solution strategies
- 🧮 Solver: Performs computations and reasoning
- ✅ Verifier: Checks consistency of reasoning
- 📝 Answer Generator: Produces clear final responses
Performance Results (GSM8K Dataset):
Multi-Agent System
61-64%
Zero-Shot Accuracy
Single Agent
16-33%
Zero-Shot Accuracy
*Meta Llama3.2-3B on Ollama is used for the evaluation.
Technologies: Python, Multi-Agent Architectures, LangGraph, LangChain, Ollama, Llama3.2-3B
Impact: Demonstrated that specialized agent collaboration significantly reduces hallucinations and improves mathematical reasoning accuracy by 2-4x compared to single-agent approaches.
Agent workflow showing the linear chain structure from Interpreter to Answer Generator
Hybrid Architecture Language Model Research (Suspended; Lack of funding)
Cutting-Edge AI Architecture Investigation
Conducted advanced research into hybrid language model architectures, exploring novel approaches to improve AI performance and efficiency through the combination of transformer and diffusion models.
Research Focus:
- 🔬 Model Architecture: Hybrid model of transformer (Left to Right) and diffusion model (Denoising)
- 🎯 Optimization: Efficiency and accuracy improvements
- 🔥 Future Adaptation: Adaptable to thermodynamic computers by transforming to energy-based models
- 🤯 Hebbian Learning: Simulates Hebbian learning through diffusion model enabling local learning rules
Technologies: Diffusion LLM, Energy-based Models
Impact: Contributes to the fundamental understanding of language model architecture and explores the future of AI computation.
Plan-and-Solve Hybrid dLLM/tLLM Architecture
Advanced CoT Reasoning with Diffusion-Transformer Hybrid System
Developed an innovative hybrid architecture that combines diffusion LLMs for high-speed planning with transformer LLMs for precise solving, achieving significant performance improvements in chain-of-thought reasoning tasks. The system leverages Mercury's superior 768 tok/sec processing speed for optimal efficiency.
Performance Benchmarks
Hybrid Architecture
384.3 token/sec
Diffusion and Transformer
Conventional Model (o3)
175.2 token/sec
RL Chain of Thought
The hybrid architecture outperforms the conventional model by 120% in token/sec.
System Architecture
# Core Implementation: Plan → Structure → Solve
async def hybrid_plan_solve_cot(user_input: str):
# Phase 1: dLLM - High-Speed Planning & Structuring
plan_result = await dllm_planner(user_input)
# Phase 2: tLLM - Final Problem Solving
final_answer = await tllm_solver(plan_result)
return final_answerTechnical Features
Speed Optimization
- Parallel Processing: Ultra-fast structuring via parallel diffusion
- Optimized Reasoning: Structured planning reduces computation overhead
- Hybrid Efficiency: Combined approach outperforms traditional single-model CoT
Error Correction Mechanism
Multi-Model Validation:
dLLM Structuring → tLLM Verification
Cross-architecture validation reduces reasoning errors through complementary model strengths
Methodology Comparison
| Traditional Methods | Hybrid Approach |
|---|---|
| Single model handles planning and solving | Specialized models for planning (dLLM) and solving (tLLM) |
| Sequential processing throughout | Parallel structuring followed by targeted solving |
| Single reasoning validation | Multi-stage verification with error mitigation |
Implementation Phases
Phase 1: dLLM Planning
- • Rapid input structuring and analysis
- • Problem categorization and strategy formulation
- • Parallel diffusion processing for pre-reasoning
Phase 2: tLLM Solving
- • Natural language processing of structured plans
- • Logical consistency verification and validation
- • Clear, comprehensive response generation
Research Validation
Theoretical Foundation:
- 1. Architectural Diversity: Different reasoning mechanisms between diffusion and transformer models
- 2. Sequential Validation: Two-stage processing reduces error propagation
- 3. Complementary Strengths: Combines fast parallel processing with precise sequential reasoning
Recent studies demonstrate significant performance improvements through heterogeneous model configurations, supporting the theoretical foundation of this hybrid approach.
Technologies: Diffusion Language Models, Transformer Language Models, Plan-and-Solve Prompting, Hybrid Architecture, Multi-Stage Reasoning, Error Correction Systems
Impact: This hybrid dLLM/tLLM architecture demonstrates the potential for specialized model collaboration to achieve superior performance in complex reasoning tasks, contributing to advancements in efficient AI system design and multi-model orchestration strategies.
Research Experience
Collective Intelligence through Multi-Agent Collaboration
University AI Research Group | Team Leader & System Architect
Led a 4-member international research team in developing a breakthrough multi-agent system that demonstrated how smaller language models can outperform large-scale models through strategic collaboration.
Key Responsibilities:
- 🏗️ System Architecture Design: Designed the overall structure with specialized agent roles
- - Prompt Generator: Break down the user input into understanable tasks for the agents
- 💻 Technical Implementation: Led the complete system development and integration
- 🌍 International Collaboration: Coordinated with international team members, conducting research discussions and presentations in English
- 📊 Performance Validation: Achieved remarkable improvement from around 70% to 90.0-92.5% accuracy, surpassing its larger sibling model (Multi-Agent GPT-4o-mini beats GPT-4o)
Technical Innovation:
- 🧠 Implemented collective intelligence principles in AI agent interactions
- 🎭 Designed role-based agent architecture for optimized collaboration
- ⚔️ Developed novel debate structures for enhanced reasoning performance
Leadership & Communication:
- 🎤 Successfully presented complex technical concepts to diverse, multilingual team
- 📜 Created comprehensive presentation materials and system documentation in English
- 🤝 Built team consensus through logical argumentation and thorough preparation
Impact: This research demonstrates the potential for efficient AI systems that achieve superior performance through intelligent collaboration rather than scale alone, contributing valuable insights to the field of multi-agent systems and collective intelligence.
Skills and Experience
Frameworks & Libraries
Django
Stripe
Vanilla JS
Bulma
OpenAI
AutoGen
LangChain
Torch
Huggingface
EvoJAX
vLLM
Ollama
React
Vite
Tailwind
Programming Languages
Python
5+ years experience
AI/ML, Backend Development, Data Science
JavaScript
3+ years experience
Frontend Development, Full-Stack Applications
HTML
3+ years experience
Web Structure, Semantic Markup
CSS
3+ years experience
Modern Styling, Responsive Design
Swift
6+ months experience
iOS Native App Development
Git/GitHub
2+ years experience
Version Control, Collaboration, CI/CD
Technical Expertise Summary
Specialization: AI/ML Systems, Multi-Agent Architectures, Full-Stack Web
Development
Core Strengths: Python-based AI development, Modern web technologies, Research-driven
development
Innovation Focus: Cutting-edge AI research, Scalable web applications, Cross-platform
development
Let's Connect
Open to Job Opportunities and Funding discussions
Loading...
Ready to Collaborate
I'm passionate about pushing the boundaries of AI and software engineering. Whether you're looking for innovative solutions, research collaboration, or technical expertise, I'm excited to discuss how we can work together to create something extraordinary.