Applied AI Engineer · LLM Agents · Industrial AI · Time-Series ML

Zhengxuan Yuan

I build applied AI systems for industrial workflows — from LLM agents and engineering knowledge reuse to time-series ML, signal validation, CAE/GNN simulation analytics, and cloud-based inference services.

kevinsyuan@gmail.com github.com/Kevoyuan linkedin.com/in/zhengxuan-yuan Munich, Germany

About

Who I Am

I’m an Applied AI Engineer based in Munich, working at the intersection of GenAI, industrial data, and engineering workflows. Since 2023, I have worked on applied AI projects in automotive and industrial engineering environments, turning ambiguous domain problems into validated AI workflows, prototypes, and practical tools.

My work spans LLM workflows, case-based reasoning, model validation, time-series and signal processing, computer vision, CAE/GNN simulation analytics, and cloud-based AI services. At IDIADA, I apply AI to crash safety and engineering workflows, from signal validation and GNN-based simulation analytics to retrieval-augmented knowledge systems.

Beyond my professional work, I build full-stack AI products such as AgentSCAD and AI-BI, where I explore agentic workflows, data applications, and user-facing AI systems. I care about modular architecture, reproducible experiments, clear evaluation logic, and AI systems that engineers can actually use.

Work Experience

Professional

Case-to-Skill Agent

LLM Agent

Co-engineered Contribution · Case-Based Reasoning

Contributed to and co-engineered a case-based LLM workflow for engineering knowledge reuse across documents, reports, and raw technical files.

~10% Accuracy Gain
~50% Inference Cost Cut
Python RAG Case-Based Reasoning AI Evaluation Cost Optimization
Highlights
  • Case-based LLM workflow for engineering knowledge
  • Skill synthesis, evaluation, fallback handling
  • Practical reliability: caching, structured outputs, cost controls

arTIco-CAE

ML

Time-Series ML

AI-powered methodology for validating crash test digital twins — combining expert knowledge with machine learning. Trained on 1,000 expert-rated simulation variants with a custom arTIco metric (-1 to +1) for qualitative model assessment.

TensorFlow sktime Optuna InceptionTime
Highlights
  • CNN, LSTM-FCN, and advanced time-series classifiers
  • Custom domain-weighted scoring metrics
  • Stratified 5-fold CV with bootstrap confidence intervals
arTIco project

Official publication on Applus+ IDIADA →

Signal Detection

ML

Binary Classification

Binary classifier system for detecting broken or abnormal measurement signals in automotive testing scenarios, with domain-specific CFC filtering.

PyTorch XGBoost SVM Transformer
Highlights
  • CFC low-pass filtering per SAE J211 standards
  • 10+ model architectures compared
  • Threshold tuning for NOK recall optimization

Crash Test Signal

Deploy

MLOps

AWS Lambda-based ML inference service for validating crash-test sensor channels, with PyTorch hybrid model and real-time signal processing.

AWS Lambda S3 SageMaker Docker ECR PyTorch
Highlights
  • ResNet18 hybrid model combining signal images with tabular features
  • Containerized Lambda deployment
  • Pydantic schemas for type-safe event handling

NavPack GNN

Research

Contributed to · Graph Neural Networks

Contributed to GNN-based crash-simulation analytics and surrogate-model workflows, including data extraction, prediction/evaluation tooling, and validation dashboards.

PyTorch NavPack PyVista HIC15
Highlights
  • Autoregressive timestep rollout prediction
  • HIC15 safety metric calculation with Numba JIT
  • Multi-trial comparison dashboards
NavPack GNN for CAE

Product page on IDIADA Digital Solutions →

EV Charger Detection

CV

Computer Vision

End-to-end computer vision pipeline for EV charger imagery — YOLO auto-labeling, ResNet50 embedding filtering, and model classification training.

YOLO ResNet50 PyTorch Selenium
Highlights
  • Human-in-the-loop annotation refinement
  • Cosine similarity filtering with pretrained embeddings
  • Leakage-aware train/val splitting by source URL
Interactive Demo
Active Learning Console
Human-in-the-loop annotation refinement
Top-K selection 0/2 batch 0 rounds
Click to explore

Signal Validation

ML

Signal Processing

Time-series signal preprocessing and validation system with domain-specific filtering, feature hashing, and cross-validation evaluation.

scikit-learn SciPy Optuna ONNX
Highlights
  • FeatureHasher for structured channel encoding
  • ONNX export for cross-framework deployment
  • Calibration and threshold analysis plots

Academic Projects

University Work

CellVision

CV

Academic Project · Computer Vision · Human-in-the-loop ML

A Streamlit-based blood-cell image analysis platform for segmentation, labeling, classification, active learning, prediction correction, and final dashboard reporting.

Demo video
Python Streamlit OpenCV scikit-learn SAM-style segmentation Active Learning
Highlights
  • Built an end-to-end cell analysis workflow from image segmentation to prediction review.
  • Added human-in-the-loop labeling and active learning to improve classifier quality.
  • Designed result and dashboard views for model evaluation and blood-cell distribution analysis.

Tour Into the Picture

CV

Academic Project · Computer Vision · 3D Reconstruction · Interactive Demo

Interactive single-image 3D reconstruction demo using vanishing-point geometry, spidery mesh generation, 5-plane room extraction, and virtual camera preview.

I implemented the geometry-focused part of the system, including vanishing-point alignment, rear-wall rectangle interaction, spidery mesh generation, 12-point room layout estimation, and a browser-based interactive demo for visualizing the reconstruction workflow.

The interactive GUI application allows users to segment foreground objects, define the vanishing point and rear-wall rectangle to extract 5 background regions, and construct a 3D scene model to generate walkthrough animations from a single 2D image.

Computer Vision Single-View Reconstruction Vanishing Point Geometry Perspective Transform Canvas / SVG JavaScript MATLAB
Highlights
  • Built an interactive browser demo for the classic Tour Into the Picture reconstruction workflow.
  • Implemented user-guided vanishing point and rear-wall rectangle alignment.
  • Generated a spidery mesh to split one image into ceiling, floor, rear wall, left wall, and right wall regions.
  • Visualized the reconstructed room as a connected 5-plane space with camera controls.
  • Adapted an academic MATLAB computer vision project into a portfolio-friendly HTML demo.
Education

Technical University of Munich

M.Sc. Mechanical Engineering

Languages
Chinese Mandarin Native
Chinese Cantonese Native
English Business fluent
German Business fluent

Personal Projects

Open Source Work

AgentSCAD

Full-Stack AI

A full-stack AI platform that translates natural-language part descriptions into parametric OpenSCAD code through a multi-stage pipeline — generation, rendering, validation, and delivery. Features real-time SSE streaming and 3D visualization.

Next.js React 19 Prisma Three.js OpenSCAD trimesh
Highlights
  • Persistent CAD job history and versioned artifacts
  • OpenSCAD render and STL artifact workflow
  • Deterministic mesh and manufacturing validation
AgentSCAD overview AgentSCAD example

Prototype data-agent pattern with controlled execution and auto-retry

AI-BI Prototype

Prototype data-agent pattern with controlled execution and auto-retry, combining business analytics, semantic routing, forecasting, and guarded code-generation workflows.

Streamlit DeepSeek Gemini Prophet SARIMA XGBoost
Highlights
  • Router Agent with semantic intent classification
  • Sandboxed code-gen data agent with auto-retry
  • Multi-model ensemble with dynamic weight optimization

ds-mon

macOS Utility

A tiny macOS menu bar app for monitoring DeepSeek API balance at a glance, with configurable refresh intervals, Keychain-backed API key storage, and customizable menu bar display.

DeepSeek Monitor menu bar dashboard
DeepSeek Monitor balance screen DeepSeek Monitor API settings DeepSeek Monitor menu bar settings
Swift SwiftUI macOS Keychain DeepSeek API
Highlights
  • Menu bar balance display with auto-refresh
  • Secure API key storage via macOS Keychain
  • Customizable icon, text, color, and refresh behavior

TradingAgents UI

Finance AI

Lightweight Streamlit interface for TradingAgents, with local app launchers, real-time multi-agent analysis progress, embedded HTML reports, report history, and optional same-Wi-Fi phone access.

Embedded report Live monitor Report viewer Report history
Python Streamlit TradingAgents CLI
Highlights
  • Live agent progress, messages, tool calls, and timing
  • Embedded HTML and Markdown report viewer
  • macOS app launcher, Windows launcher, and LAN mode