Zhongyuan Wang

Founding ML Scientist · RaptorX.AI, Remote2024 - Present

Built a production-style Agentic Engineering workflow for bank fraud alerts: Agent Loop, context management, planning, Guardrails, case-store state, Evidence/Report agents, SFT/RLVR/RLHF, evaluation, and runtime/MLOps.
Built criteria-driven multi-agent systems for paper/patent production; as lead of a five-person ML team, owned technical direction, experiment checks, drafting, counsel handoff, and audit logs.
Built the underlying unsupervised fraud engine from zero, combining GNNs, behavioral autoencoders, explainability, and an edge-probability model across 30 relationship types to reduce dependency on fraud labels.
Achieved 328× candidate reduction (590K to 1,883 high-confidence alerts) at 99.9% fraud-type specificity.
Converted PoCs into 5 signed enterprise customers across the US, India, and the Gulf: Payactiv (US), NSDL and Lenskart (India), Omantel and Thawani Pay (Oman); US deployment covers 460K+ users across 6,700+ employers.

Senior Engineer, AI & Data Platform for Autonomous Driving · Lotus Tech (Geely Group), Hangzhou2021 - 2024

Led an in-house video anonymization pipeline from vehicle data ingestion to data-center processing, supporting 80+ TB/day fleet-scale autonomous-driving data.
Trained and deployed YOLOv5-based face/license-plate detection across 10M+ frames, achieving 90%+ recall for privacy-sensitive objects.
Built a Volcano-based Kubernetes GPU job scheduling platform for PyTorch workloads on a local 64-GPU NVIDIA A100 80GB cluster (8 nodes × 8 GPUs), improving utilization and saving roughly RMB 4M/year versus Alibaba Cloud.
Optimized GPU-accelerated batch inference for fleet-scale video processing, reducing latency 30% and improving throughput.
Replaced third-party vendors with an internal ML platform, saving roughly $1M/year while supporting GDPR/PIPL compliance.

Independent ML Study & Graph-ML Projects2020 - 2021

Self-directed study of graph machine learning: implemented GNN models for node classification and link prediction on public datasets.

Software Development Engineer, Financial Anti-Fraud · DataVisor, Shanghai2020

Built an unsupervised loan application fraud model combining GNN and traditional ML, using hundreds of features across behavioral, device, third-party credit, and raw transaction data.
Directly contributed to securing a $500,000 enterprise contract through technical leadership and client demonstration.

Research Intern, DiDi AI Labs · UCL-DiDi Master Research Program, Beijing2019

Leveraging Graph Neural Networks for user travel intent prediction, supervised by Prof. Jieping Ye (VP of DiDi Chuxing, IEEE Fellow).
Selected as 1 of only 2 projects (out of 30+ candidates) across the entire UCL cohort for the DiDi AI Labs research program.

Publications

Patents

Inventor on 20 applications spanning graph-based fraud detection and LLM/agent systems: 6 filed (US), 4 published (India), 10 in draft.

Published (India) App Nos.: 202541024623, 202541045354, 202441079562, 202541032797

Experience