Building AI agents, music intelligence systems, and adaptive learning experiences.

I'm a high school student and developer exploring the intersection of AI Agents, Music Technology, and Education.

I am especially interested in how Large Language Models can move beyond simple question-answering systems and become adaptive reasoning architectures capable of perception, planning, action, feedback, and long-term personalization.

My projects focus on turning these ideas into real, interactive products — from AI-powered choir training and adaptive education agents to music-driven games and multi-agent reasoning systems.

AI-powered, part-specific choir practice and creative learning platform

PartMate-MTX is a browser-based practice system built specifically for choir sectional rehearsal. Instead of reducing a performance to a single entertainment-style score, PartMate creates a complete learning loop: locate the mistake → diagnose why it happened → prescribe a targeted exercise → verify mastery through an unguided retest.

During practice, the system uses MusicXML scores and teacher reference recordings as musical targets, analyzing pitch, entry timing, missed notes, rhythm, note duration, and stability down to individual phrases and measures. Pitchy / MPM provides pitch tracking, while RMS energy and spectral flux provide complementary evidence for onset and missed-note detection. Thresholds and scoring rules are centralized in a versioned scoring dictionary, keeping real-time feedback, review, prescriptions, and retests consistent. Deterministic rules make the musical judgments; the LLM does not score performances, but translates structured results into understandable feedback and encouragement.

Beyond scoring, PartMate integrates OMR score digitization, OSMD-based MusicXML rendering, Web Audio synchronization, and Chroma + DTW alignment between teacher demonstrations and sheet music. A personalized growth system tracks performance by song × vocal part × measure, dynamically estimates mastery, and gradually removes guidance before unguided retesting. The conductor dashboard aggregates individual results into a part × measure heatmap, helping identify shared weaknesses before rehearsal.

The Create / Harmony Workshop extends PartMate from practice into musical creation. MVSep separates vocals from accompaniment, pYIN extracts melodic information and supports key, chord, and beat analysis; rule-based logic preserves musical constraints while the LLM directs section-level and vocal-part arrangement. Pitch shifting, SVC voice conversion, and multi-track mixing then transform a single vocal line into editable layered harmonies.

The goal is to move users from learning a song → building personalized practice habits → understanding ensemble relationships → arranging, harmonizing, and creating music themselves.

Core Technologies: TypeScript · MusicXML / OSMD · Web Audio · OMR · Pitchy / MPM · RMS & Spectral Flux · Rhythm Analysis · Chroma + DTW · Rule-based Scoring · LLM Feedback · MVSep · pYIN · SVC · Multi-track Audio

Beat-Runner is an original pseudo-3D web rhythm runner built around music-driven interaction. Players control an energy ball across three neon tracks, dodging obstacles and collecting notes in sync with each song's BPM. The game includes three levels, rhythm-based abilities, talent choices, combo and ranking systems, story progression, level unlocking, and responsive controls for both desktop and mobile devices.

Tech: JavaScript · HTML5 · CSS3 · Web Audio · BPM Synchronization

Pathfinder-AI-Learning is an adaptive education agent that explores how LLMs can support individualized learning rather than simply answer questions. It combines agentic planning, memory-decay modeling, dynamic knowledge graphs, and motivational feedback to transform static learning materials into a responsive learning process. The project focuses on long-term adaptation, personalized pacing, and incentive design for sustained student engagement.

Tech: LLM · AI Agents · Knowledge Graphs · Mathematical Modeling · Adaptive Learning

Terminal_Detective is an experimental AI reasoning game in which the player acts as the architect behind the detective. Instead of solving cases directly, players design ReAct-style behavior logic that guides an AI detective through observation, reasoning, tool use, and recovery from mistakes. The project explores multi-agent coordination, controllable reasoning workflows, and fault-tolerant agent design through an interactive game format.

Tech: ReAct · Multi-Agent Systems · LLM Reasoning · Agent Workflow Design

• AI Agent Architecture — Perception → Planning → Action → Feedback
• AI + Music — Music understanding, rehearsal intelligence, and creative tools
• Adaptive Education — Personalized learning systems and long-term learner modeling
• Multimodal AI — Combining audio, vision, structured data, and language models
• Human-AI Interaction — Designing AI systems that adapt to how people actually learn and create

AI & Agents

LLM · Prompt Engineering · Agentic Workflows · ReAct · Multi-Agent Systems

Music & Signal Processing

Audio Analysis · Pitch Detection · Rhythm Analysis · OMR · Multimodal Processing

Web Development

JavaScript · TypeScript · HTML5 · CSS3

Research & Modeling

Mathematical Modeling · Knowledge Graphs · Behavioral Economics · Adaptive Systems

I'm interested in building AI systems that do more than generate answers.

I want to explore systems that can:

understand users → model their progress → make decisions → adapt over time → help them learn and create better.

My long-term interest lies in the intersection of AI Agents, intelligent education, music technology, and human-centered AI.

Email: chenyuhang987@icloud.com

Phone number: +86 13146900226