Hi, I'm Jiayi Lu. I turn data into decisions, models, and the occasional aha.

P / 01

SGT-Aug · Semantics-Guided Augmentation for Long-Tailed Multi-Label Classification · ICME 2026 (CCF-B, 28.89% acceptance) · 1st author · Accepted, EI indexed by Nov

Inject a pretrained Transformer to model label-wise semantic relations, enabling semantic–vision interaction with a ResNet50 backbone to localize tail-class regions and synthesize targeted augmentations that ease tail-class sparsity.

Problem: in long-tailed multi-label image classification, tail classes are sparse and unevenly distributed — naive training tends to "squeeze them flat".

Approach: ResNet50 backbone + pretrained Transformer for semantic–vision interaction, localizing tail-class-relevant foreground regions. Augmentation is built on those localizations to ease the long-tail imbalance without amplifying it.

Training & results: Linux + PyTorch with single-node multi-GPU DataParallel. Systematic comparison and ablations on VOC-LT / COCO-LT — tail-class accuracy improves by 5.55 / 10.33 percentage points over comparison baselines, and combined with standard classification losses it outperforms several long-tailed multi-label methods.

• PyTorch
• ResNet50
• Transformer
• DataParallel
• VOC-LT / COCO-LT

P / 02

SlayQuiz · JD-Driven AI Interview-Training Mini-Program

For job-seekers drowning in dense JDs and fragmented prep paths — decomposes unstructured JDs into role profiles, skill graphs, and tiered training paths, closing the loop from "understand → practice → feedback → review → improve".

System design: sole owner of architecture, technical selection, and prototyping. Decoupled AI generation from deterministic rules (scoring / XP / unlock / stats) and built the closed loop around "JD-driven, personalized interview training".

Stack: Taro + TypeScript mini-program (with H5 cross-end extension) on the front; FastAPI on the back for jobs, question sets, attempts, reports, auth, and KB endpoints, synced via request/response contracts.

AI layer (LangChain + DeepSeek): LangChain orchestrates DeepSeek for JD analysis and on-demand question generation; Bocha / Tavily web search enriches the JD context. JSON Mode + Pydantic Schema parse model output, while question-type quotas, answer-cardinality, knowledge-point binding, and question de-duplication gate non-compliant content.

Generation reliability: async task + fault-tolerance pipeline — status polling, idempotency keys, bounded retries, and tiered recovery handle model timeouts, network errors, JSON parse failures, and rule-check failures. Repeated or malformed questions are repaired in place, avoiding full-set regeneration. Scoring, XP, unlock, and stats are computed by deterministic code; uncertainty is contained at the content-generation step.

Hybrid RAG + learning loop: PDF / Word / URL ingestion, chunking, DashScope Embedding, and Qdrant retrieval with per-user data isolation. Learning layer combines mistake notebooks, variant questions, staged spaced repetition (1/3/7/14 days), and adaptive difficulty across basic / applied / interview tiers — driven by the last-20 accuracy, sample size, and mistake knowledge points — with XP rewards to shape the training path.

Key results: question-generation latency P90 < 15s; in-round semantic duplication 0, cross-round semantic duplication < 5%; full closed loop from "JD understanding → training plan → on-demand generation → instant feedback → mistake review → learning report", with multi-job support and per-user KB isolation.

• Taro · TypeScript
• FastAPI
• LangChain · DeepSeek
• Hybrid RAG · Qdrant
• Pydantic · JSON Schema
• Async Tasks · Retry

P / 03

Strategem · Multi-Agent Decision-Intelligence Platform

For energy-storage / virtual-power-plant industry research — turning "reading reports" into a runnable pipeline where multi-role agents collaborate to produce traceable decision intelligence.

Deep Research Agent, end-to-end: built on FastAPI + LangGraph + Next.js — supports research-plan confirmation, report generation, section-level rewriting, SSE streaming of task-stage progress, and reference-evidence inspection.

State machine + routing: LangGraph state machine + ResearchState orchestrates intent recognition, plan generation, retrieval, evidence assembly, analysis, and writing. Routing combines "keyword-rule fallback + LLM structured classification" for task triage, low-confidence clarification, and plan confirmation — 96% routing accuracy verified by offline evaluation.

Evidence-coverage driven loop: research flow is controlled by evidence coverage and an iteration budget — when uncovered sub-questions remain and the iteration cap is not yet hit, a supplementary query is generated and fed back to retrieval; otherwise the report is produced. Existing evidence chains are reused for section-level rewrites after report completion.

Dual retrieval + evidence pool: "web search + PostgreSQL + pgvector local KB" dual-channel pipeline; multi-format upload → parse → chunk → DashScope Embedding → pgvector recall. Tenant- and user-scoped data isolation; web sources are ranked by semantic relevance, authority, and freshness, and two-level de-duplication (URL/title + content hash) forms a traceable unified evidence pool.

Evidence chain + audit: four-level "source → evidence → conclusion → citation" chain and audit layer — source whitelist, two-level de-duplication, back-end citation derivation, and structured-output constraints filter illegal or unsupported claims. Conflicting data, missing modules, and structured-output failures are audited, retried, or downgraded; uncovered items are written to the pending-verification queue.

Ownership & outcomes: led the technical design and core pipeline of the Deep Research Agent — LangGraph state machine, ResearchState, dual retrieval + evidence pool, coverage-driven supplementary search, citation derivation + audit markers, and structured-output retry/downgrade. The system supports multi-tenant isolation, section-level rewrites, and a fully traceable citation graph.

• LangGraph · State Machine
• Deep Research Agent
• Next.js
• FastAPI · SSE
• Postgres + pgvector
• Evidence Binding

P / 04

Ensemble Learning for Time-Series Forecasting · CPCI · 1st author

Financial time series are highly nonlinear and noisy — a single LSTM is unstable. The paper boosts robustness with a Bagging + LSTM ensemble.

Feature engineering: EDA on historical stock data to surface open / close / volume signals and craft multi-dimensional features; Fourier and B-spline basis functions for smoothing and feature mapping.

Core model: TensorFlow-based LSTM capturing long-range dependencies.

Ensemble & validation: Bagging constructs diversified training subsets, training multiple LSTMs in parallel. Across multiple forecast horizons, prediction error is significantly lower than the single-LSTM baseline.

• TensorFlow
• LSTM
• Bagging
• Fourier / B-spline
• Time-Series

P / 05

Chest X-Ray Intelligent Recognition & Statistical Modeling · National Statistical Modeling Contest · Team Lead

CNN for local texture + GNN for inter-region relations — pneumonia classification on chest X-rays.

Problem: lesion regions are scattered; local-texture-only models miss spatial-relation cues.

Approach: build a region adjacency graph via SLICO superpixel segmentation, extract per-node features with CNN, then model inter-region relations with GNN — a joint local + relational representation.

Results: systematic experiments & evaluation lifted headline metrics to 92% accuracy, 91.5% sensitivity, 99.5% AUC; as team lead, owned the technical roadmap, experimental design, and coordination.

• PyTorch
• CNN
• GNN
• SLICO Superpixel
• Medical Imaging

P / 06

Diabetes Prediction via a Stacking-Ensemble Framework · PIMA · Stacking Ensemble

Stacking LR / SVM / CART with AdaBoost / GBDT — pushed PIMA AUC to 0.96.

Preprocessing: KNN imputation for missing values, then random undersampling to address class imbalance; 80/20 train-test split, with numerical features normalized.

Base learners + boosting: Logistic Regression, SVM, and CART as base classifiers; CART further wrapped into AdaBoost and Gradient Boosting ensembles.

Stacking + hyperparam search: 5-fold cross-validation with Bayesian optimization (Optuna). The fused Stacking model hit AUC 0.96 / Recall 0.84 / Accuracy 0.89 — a 10.3-point AUC lift over the logistic-regression baseline.

• Logistic Regression · SVM
• CART Decision Tree
• AdaBoost · GBDT
• Stacking
• Optuna
• PIMA