Independent AI Researcher · Machine Learning Systems Developer · Robotics Engineer

Exploring speech intelligence, multimodal reasoning, grounded language models,
and embodied AI systems for real-world human–machine interaction.

I am an independent high school researcher and AI systems developer based in Taiwan, working at the intersection of Machine Learning, Speech Processing, Large Language Models, and Intelligent Robotics.

My work focuses on transforming research concepts into complete and experimentally verifiable systems. Rather than treating an AI model as an isolated component, I study how representation learning, knowledge retrieval, inference infrastructure, evaluation, human interaction, and physical embodiment can be integrated into a coherent architecture.

My primary research interests include:

• Speech-native language understanding
• Low-resource language technology
• Prosody-aware semantic reasoning
• Audio-text multimodal learning
• Retrieval-Augmented Generation
• Evidence-grounded language models
• Explainable and trustworthy AI
• Human–robot interaction
• Embodied intelligence
• Local and distributed model inference

I am particularly interested in AI systems that move beyond surface-level language processing to reason about intent, evidence, uncertainty, cultural context, and real-world constraints.

Conventional speech systems commonly decompose intelligence into a sequential pipeline:

Speech → Automatic Speech Recognition → Text → Language Model

Although effective, this architecture may discard acoustic and paralinguistic information during transcription.

My research explores whether models can reason directly from speech representations while preserving information such as:

• Prosody and intonation
• Emotion and speaking style
• Pragmatic intent
• Speaker relationships
• Regional variation
• Cultural and situational context

The long-term objective is to develop speech systems capable of jointly modeling acoustic, linguistic, and contextual information rather than depending exclusively on textual intermediates.

I am investigating machine learning methods for languages with limited labeled resources, with particular interest in Taiwanese Hokkien.

Current areas of exploration include:

• Speech representation learning
• Cross-lingual transfer learning
• Audio-text joint instruction tuning
• Low-resource ASR evaluation
• Dialect-aware retrieval
• Prosody-aware understanding
• Authorized voice preservation
• Speech dataset design and governance

This direction examines not only model accuracy, but also data provenance, linguistic variation, evaluation validity, speaker consent, and responsible deployment.

Large Language Models can generate fluent responses while still producing conclusions that are unsupported, incomplete, or contextually inappropriate.

I design grounded AI pipelines that combine:

Structured Knowledge
        ↓
Hybrid Retrieval
        ↓
Metadata and Context Filtering
        ↓
Neural Reranking
        ↓
Grounded Generation
        ↓
Evidence and Confidence Verification

The objective is not merely to generate an answer, but to construct systems capable of explaining:

1. What evidence was retrieved
2. Why the evidence is relevant
3. Which claims are supported
4. How uncertainty is represented
5. When the model should abstain
6. When human verification is required

My robotics work examines how machine intelligence can be connected to physical systems through perception, reasoning, decision-making, and action.

A complete embodied AI system may integrate:

• Speech and visual perception
• Environmental sensing
• Multimodal representation learning
• Language-model reasoning
• Structured decision policies
• Embedded controllers
• Motors and actuators
• Displays, lighting, and audio feedback

The central research question is how an intelligent system can convert incomplete and uncertain observations into safe, interpretable, and useful physical behavior.

Context-Aware AI for Cultural Knowledge Preservation and Communication

CultureBridge is a multimodal AI and robotics system designed to address cultural misunderstanding caused by missing social and contextual information.

Unlike conventional translation systems, CultureBridge analyzes not only literal language, but also:

• Tone
• Forms of address
• Interpersonal relationships
• Age and hierarchy
• Regional conventions
• Cultural taboos
• Formality
• Interaction context

Cultural Records
        ↓
Metadata-Aware Retrieval
        ↓
Evidence Reranking
        ↓
Contextual Reasoning
        ↓
Confidence and Citation Verification
        ↓
Speech and Robotic Interaction

• Cultural knowledge database
• Hybrid Retrieval-Augmented Generation
• Region and dialect filtering
• Automatic Speech Recognition
• Prosody-aware analysis
• Text-to-Speech synthesis
• Evidence and confidence interfaces
• Human review workflow
• Interactive physical robot

The system follows a central principle:

AI should not independently define culture.

Instead, the model should retrieve documented knowledge, expose supporting evidence, represent uncertainty, and defer to human expertise when available evidence is insufficient.

Technical Areas

Speech AI · RAG · LLM · ASR · TTS · FastAPI · Robotics · Human–AI Interaction

Deep Reinforcement Learning for Autonomous Sequential Decision-Making

STDA-RL-Snake is a reinforcement learning environment in which an autonomous agent learns to play Snake through interaction rather than explicit rule-based control.

The project investigates:

• State-space representation
• Reward-function design
• Exploration and exploitation
• Temporal-difference learning
• Experience replay
• Target-network stabilization
• Policy evaluation
• Training convergence

The agent is trained using Deep Q-Learning, connecting discrete decision problems with neural value-function approximation.

Technical Areas

Python · PyTorch · Deep Q-Network · Reinforcement Learning

• Deep neural networks
• Representation learning
• Reinforcement learning
• Natural Language Processing
• Speech processing
• Multimodal learning
• Model evaluation
• Inference optimization
• Dataset construction
• Experimental analysis

• Large Language Models
• Retrieval-Augmented Generation
• Automatic Speech Recognition
• Text-to-Speech
• Voice cloning
• Speech representation models
• Audio-text joint modeling
• Evidence-grounded generation
• Context and prompt architecture

• Local LLM deployment
• Multi-GPU inference
• Linux server administration
• CUDA-based workloads
• vLLM model serving
• Hugging Face Transformers
• REST API development
• Containerized deployment
• SSH remote development
• Model-service integration

• Backend system architecture
• API contract design
• Web application development
• Data-pipeline integration
• Testing and validation
• Technical documentation
• Reproducible deployment
• Human–AI interface design

• Sensor integration
• Embedded control
• Autonomous decision systems
• Motor and actuator control
• Human–robot interaction
• Edge AI integration
• Robot–server communication
• Physical system prototyping

Python · C++ · JavaScript · TypeScript · HTML · CSS · Bash

PyTorch · FastAPI · React · Next.js · Node.js · Docker · Linux · Git

My research and engineering development has been shaped by participation in robotics competitions and international technical programs, including:

• World Robot Olympiad — WRO
• Borneo International Robot Tournament — Borneo-IRT
• FIRST Robotics Competition — FRC
• FIRST LEGO League — FLL
• Asia-Pacific Robotics Championship — APRA
• Global Robotics Games — GRG

These environments provide practical testbeds for studying intelligent systems under real operational constraints, including:

• Incomplete observations
• Hardware uncertainty
• Limited computation
• Strict time constraints
• Dynamic environments
• High-stakes decision-making
• Human–robot coordination

A convincing response is not necessarily a correct response. Intelligent systems should expose the evidence supporting their conclusions.

A model demonstration is not a complete system. Practical intelligence requires data pipelines, interfaces, evaluation, deployment, monitoring, and failure handling.

A trustworthy model must be able to express uncertainty, abstain from unsupported conclusions, and request human verification.

Human communication depends on social, cultural, acoustic, and situational information that cannot always be represented by text alone.

Technical claims should be supported by documented configurations, measurable evaluation criteria, and repeatable experiments.

My long-term objective is to develop intelligent systems capable of understanding speech, reasoning with evidence, recognizing uncertainty, and interacting reliably with people and physical environments.

I am especially interested in systems that unify:

Perception
    +
Representation Learning
    +
Knowledge Retrieval
    +
Reasoning
    +
Decision-Making
    +
Physical Interaction

The goal is not simply to construct increasingly large models, but to develop AI systems that are:

• Grounded in verifiable evidence
• Interpretable in their reasoning
• Aware of uncertainty
• Sensitive to linguistic and cultural context
• Experimentally reproducible
• Operationally useful in the real world

From machine learning models to complete intelligent systems.