Research & Engineering

李晨悦

Research Engineer / AI Engineer

LLM · Multimodal · Agent · Evaluation

I am an undergraduate at Jiangnan University, expected to graduate in 2028. My projects span agent execution and failure attribution, recommender evaluation, and multimodal data selection and fine-tuning, with a focus on reliable AI systems and reproducible experiments.

CV GitHubGoogle ScholarEmail

Academic Research

All academic research

Independent research, engineering explorations, and disclosure-safe updates, organized around questions, methods, and evidence.

Swipe horizontally to inspect the full figure

MiniClaudeCode: Ordered Trajectory Credit & Agent Evaluation

01 / Independent project Offline estimator study

Research question: How should actions receive credit when tool trajectories contain prerequisites, interactions, and failure recovery?
Method: Constrain counterfactual replay with a precedence DAG, combining exact enumeration, uniform sampling of valid orders, and offline evaluation.
Key results: Under the controlled protocol, precedence-constrained estimates are closer to the prefix reference than vanilla Shapley.
Exact estimation, sampling convergence, and replay-validity analyses are publicly documented.

Code Project page

Swipe horizontally to inspect the full figure

KAI Offline RecSys Lab: Retrieval-to-Reranking Evaluation

02 / Independent project Public data · offline experiments

Research question: How can retrieval, negative sampling, and reranking be compared under the same population, data splits, and candidate sets?
Method: Freeze Two-Tower retrieval and Exact / HNSW candidates, compare DIN / DCN rerankers, and apply train-only calibration with paired user-level evaluation.
Key results: DIN reranking improves ordering within the same candidate sets under the frozen offline protocol.
Uniform negatives win this dev comparison; feature dependencies and ANN trade-offs remain visible.

Code Demo Project page

Swipe horizontally to inspect the full figure

Qwen2.5-VL: Data Selection Under a Small Fine-Tuning Budget

03 / Independent project Offline experiment · public summary

Research question: With the model, training budget, and held-out evaluation fixed, can data selection improve multimodal fine-tuning over random sampling?
Method: Compare Random-1K and faithful COINCIDE-1K on ScienceQA with the same Qwen2.5-VL-3B, LoRA / SFT settings, and paired seeds.
Key results: In the public summary, COINCIDE-1K trails Random-1K across all three paired seeds.
Negative results, a confidence interval crossing zero, and selection-mismatch analysis are retained.

Demo Project page

Research & Industry

Experience

2026

Tencent Music

Java Backend & AIGC Application Engineering Intern

Dates to be added

Schneider Electric

Intern

Open Source

All contributions

Contributions, maintained repositories, and infrastructure for reproducible research.

Pull requestSubmitted · not merged (2026-09-09)

PyRIT: Garak API-key Scenario & Scorer

microsoft/PyRIT

External contribution PR #2538 adds API-key generation and completion probes to the Garak scenario family.

Contribution: Submitted the scenario, fixed corpora, CredentialLeakScorer extensions, tests, and documentation. The implementation distinguishes credential-shaped values from public identifiers and excludes request echoes and supplied test fragments.
Result: The code diff and test notes are public. The contribution is submitted but unmerged; submission alone does not establish maintainer endorsement or validation against live targets.

Pull request Repository

RepositoryPersonal repository · source available

MiniClaudeCode: Observable Execution & Evaluation Infrastructure

hey-Chloe/MiniClaudeCode

A personally maintained coding-agent reference implementation focused on execution boundaries, tool events, checkpoints, and evaluation.

Contribution: Published a bounded control loop, tool registration and policy interfaces, session persistence, trajectory attribution, and evaluation scripts. Third-party training frameworks carry separate attribution.
Result: Source, controlled reports, and test entry points are available. Offline harness validation is distinct from real-model task solving; the repository currently grants no additional open-source license.

Repository Documentation

InfrastructurePersonal repository · research infrastructure

KAI RecSys Lab: Experiment Protocols & Reporting Tools

hey-Chloe/KAI-Offline-RecSys-Lab

Provenance ledgers, frozen configurations, evaluation reports, and portable experiment views built around public data.

Contribution: Connect retrieval, reranking, calibration, and paired statistics under one offline protocol, with synthetic-path reproduction, public-report checks, and a static Playground.
Result: Reports distinguish public-data experiments, synthetic integration checks, ANN systems measurements, and local serving rehearsals, without presenting offline findings as online business gains.

Repository Documentation

Engineering

All projects

2026Agent UI / Vue / FastAPI / SSE

MiniClaudeCode Studio

A Chinese-language agent workspace that brings plans, tool calls, approvals, code diffs, tests, and checkpoint recovery into one traceable workflow.

Problem: Make agent execution, file changes, and verification outcomes observable and distinct.
Ownership: The personal repository's Studio adapter and interface reuse MiniClaudeCode's native execution engine. The public version replays recordings; the local version connects to the actual runtime.
Outcome: A local workspace, static replays, and a small online evaluation distinguish execution ending, functional correctness, and final verification, while retaining failures caused by permissions and interpreter selection.

Code Demo Project page

2026Static analysis / Semgrep / FastAPI / Vue

RuleForge-SAST

A static-analysis workflow for PHP, Python, and Java: rule scanning, vulnerability explanations, AI repair suggestions, rescanning, and history.

Problem: Separate an AI repair suggestion from scanner findings, with inspectable evidence before and after a change.
Ownership: The personal repository connects language and rule selection, scan records, AI suggestions, and Patch Verify across the frontend and backend. Detection uses Semgrep and existing rules.
Outcome: Scanning, AI suggestions, and repair checks form a traceable workflow with a demo and scan history. FIXED means the current rule no longer matches; coverage depends on the rules used.

Code Demo

Research Notes

All writing

2026-09-12

研究进行中｜AI Coding Benchmark 可复现性审计

检查公开数据、任务资产与评测程序在不同发布视图之间是否保持一致，并建立可追溯的离线验证流程。

Research Update · AI Coding · Agent Evaluation · Reproducibility · 2 min read