AI Agent / Applied AI Engineer · ML Research
I build agentic AI systems, RAG backends, long-term memory, multimodal applications, and reproducible ML research pipelines.
My focus is not just getting a demo to run. I care about the engineering around it: tool orchestration, memory, retrieval quality, evaluation, data leakage, reproducibility, privacy boundaries, and failure analysis.
2027 graduate · Artificial Intelligence major
Target roles: AI Agent Engineer · AI Application Algorithm Engineer · Applied AI / ML Engineer
| Project | What I built | What it demonstrates |
|---|---|---|
| Multimodal Desktop Agent | Agent orchestration core with pluggable LLM providers, MCP-style tool namespaces, bounded tool loops, memory backends, voice/avatar adapters and offline tests | Agent engineering · tool use · memory · modular architecture |
| Human-Gated Multi-Agent Supervisor | Local-first agent runtime with leases, heartbeats, pending spawn approval, bounded child capacity and resumable generations | Multi-agent orchestration · state machines · failure containment |
| Medical RAG Assistant | FastAPI RAG backend with bounded document ingestion, Chroma retrieval, deterministic routing, optional read-only Neo4j context and SSE streaming | RAG · AI backend · retrieval · API engineering |
| Rhythm Chart Generation Research | Structured sequence-modeling research with deterministic legality checks, causal lane-recency priors, multi-head diagnostics and group-disjoint OOF evaluation | PyTorch · sequence modeling · leakage-aware evaluation |
| Agentic Research Control Plane | SHA-bound candidate promotion, safe state IDs, Lead gates, prospective protocol/runner/auditor freezing and fail-closed audit promotion | Agent research infrastructure · reproducibility · evaluation governance |
| OpenAI-Compatible MCP Bridge | Streamable HTTP MCP adapter for OpenAI-compatible endpoints with validation, mock upstream and real MCP E2E test | MCP · model adapters · integration testing |
| Aegis Financial Intelligence | Local-first event intelligence pipeline with rule fallback, optional LLM enrichment, SQLite audit trail, CLI and dashboard | LLM applications · structured reasoning · auditable pipelines |
| If you are hiring for | Open first | Then |
|---|---|---|
| AI Agent / LLM Application | Multimodal Desktop Agent | Human-Gated Supervisor · MCP Bridge |
| RAG / AI Backend | Medical RAG Assistant | Multimodal Desktop Agent |
| Agent Infrastructure / Evaluation | Human-Gated Supervisor | Agentic Research Control Plane |
| Applied ML / Algorithm | Rhythm Chart Generation Research | Agentic Research Control Plane |
- Multimodal Desktop Agent: 5/5 public tests pass; ruff clean.
- Human-Gated Multi-Agent Supervisor: 10/10 state/coordination tests pass; ruff clean.
- Medical RAG Assistant: 17/17 tests pass; ruff clean.
- Rhythm Chart Generation Research: 7/7 public tests pass; ruff clean.
- Agentic Research Control Plane: 8/8 promotion/control tests pass; ruff clean.
- OpenAI-Compatible MCP Bridge: 3 validation tests + offline smoke test + real MCP client/server E2E pass; ruff clean.
Some of my largest systems stay private because they contain local datasets, copyrighted media references, user-derived evaluation artifacts, or machine-specific integrations. I publish the engineering ideas without exposing material that should remain private.
A research platform for coordinating multiple AI research workers under reproducible experiment governance.
- isolated Scientist workspaces and task ownership;
- Lead promotion gates for research results;
- protocol / runner / auditor freezing before formal experiments;
- artifact hashing and traceable handoffs;
- deterministic independent auditing before downstream promotion;
- explicit separation between research, paper validation and live execution.
Signal: multi-agent orchestration, research infrastructure, reproducibility, fail-closed evaluation.
A model-independent persistent identity and memory prototype for long-running AI interaction.
- event-sourced identity history and traceable long-term memories;
- human-confirmed memory promotion and versioned persona evolution;
- provider-independent model switching;
- portable identity export / restore;
- offline regression tests and external long-term-memory evaluation;
- explicit separation between initial persona, observed events and retrieved memories.
Signal: stateful agents, long-term memory, evaluation, provenance and local-first system design.
The full private research workspace behind the public rhythm-chart showcase.
- audio/MIR evidence extraction and temporal opportunity detection;
- typed gesture and lane-state modeling;
- long-note duration / release scheduling;
- difficulty-aware generation and simplification;
- replay, playability and human-reference diagnostics;
- preregistration, ablations and falsification-oriented experiments.
Signal: research depth, structured sequence modeling, experimental discipline.
More detail: Private Research Highlights
- Agent systems: provider abstraction, tool orchestration, bounded execution loops, memory adapters, MCP-style integrations
- RAG / retrieval: document ingestion, vector search, deterministic routing, evidence-aware synthesis, graph context
- ML research: PyTorch, sequence modeling, CV, RL, leakage-aware splits, grouped OOF evaluation
- Backend / systems: FastAPI, SQLite, REST/SSE, CLI tooling, local-first applications
- Evaluation: frozen protocols, deterministic checks, negative-result retention, reproducible experiment structure
- Foundations: Python, C/C++, data structures & algorithms
- Semantic Movie Finder — multilingual dense retrieval with SentenceTransformers + FAISS
- CLIP + LSTM Image Captioning — image captioning with CLIP features, LSTM decoding and beam search
- Flood Road Segmentation — DeepLabV3+ segmentation pipeline with deterministic data pairing
- EV Wireless-Charging DQN — DQN environment with graph action masking and deterministic tests
- 16th Lanqiao Cup, Jiangsu Division — First Prize, C/C++ Programming, University Group B (2025)
- 2025 AIComp Provincial Selection — Third Prize, algorithm challenge on predicting/recommending users' potential book borrowing
I prefer testable systems over impressive-looking demos:
- keep evaluation boundaries explicit;
- separate deterministic rules from learned components;
- preserve negative and inconclusive results;
- make private/public boundaries intentional;
- turn experiments into reproducible workflows rather than one-off notebooks.
Current focus: AI agents, RAG, persistent memory, multimodal systems, and research-grade evaluation.

