-- Published in 2025, the peer-reviewed study Real-time Industrial Surface Defect Detection Technology Based on a Lightweight Transformer is co-authored by AI engineer Wenke Du, who is listed as the third author. The paper is indexed in the IET Digital Library and IEEE Xplore and has been cited by other researchers, including one study on lightweight vision transformers for low-quality infrastructure images. It presents a lightweight Transformer model for industrial surface-defect detection and marks one throughline in Du's work: making AI dependable enough to trust in real-world settings.
The paper tackles a practical reliability problem. Industrial inspection must run in real time on imperfect images and modest hardware, conditions in which heavyweight models may be difficult to deploy. The paper presents a Transformer-based detector designed to retain detection accuracy while operating quickly enough for factory-floor use. Its abstract reports 95.8% detection accuracy and an inference speed of 45 frames per second on industrial-grade hardware. It also reports a 40% reduction in computational overhead while maintaining comparable accuracy.
Du was also a founder before her current work on reducing hallucinations in AI agents. She co-founded VIRA, an AI-enabled learning platform that reached five hundred daily users, was shown at the United Nations' Youth2030 exhibition, and won full scholarships from the Watson Institute Accelerator and Global Citizen Year; a second venture, a blockchain simulation platform, drew twenty thousand beta users and $30,000 in funding. Her early work focused on tools that other people could use and check.
That instinct now drives her engineering. Models can invent facts, predictive systems can be badly calibrated, and autonomous agents can act on unchecked conclusions. In settings involving document analysis, decision support, or automated review, an unverified result can create practical risks for users and organizations. Du's work focuses on trustworthy, verifiable AI: systems that verify their outputs, quantify uncertainty, and reduce hallucination before anyone acts.
Rather than argue in the abstract, she measures it and releases the code. Auditing real smart contracts, she built a multi-stage agent that understood the code, planned an exploit path, classified the flaw, and then re-checked itself in a read-only sandbox, cutting its hallucinated conclusions by about a fifth. In separate work on coding agents, a framework of more than forty specialized agents under a self-verifying review layer cut post-execution failures from 38% to 14%.
Her open reproductions make the mechanism checkable by anyone. On a benchmark of 121 real smart contracts, asking a model "are you sure?" barely helped, and self-judging discarded correct findings, while grounding each claim in the source drove fabricated conclusions to zero. On open intensive-care data, honest calibration cut false alarms by about a fifth at matched sensitivity; on a century of market data, a standard risk measure breached its own limit more than twice as often as promised.
What ties this together is a systems mindset focused less on any single model than on the machinery around it: how a decision is logged, replayed, attributed to its evidence, and stress-tested before it is trusted. One provenance tool she built traces a figure to the exact characters in the document it came from; another turns an agent's audit into a trail an outsider can follow.
From a virtual classroom to an AI that checks itself, the mission has not really changed. She is still building tools meant to be shared, and still asking the question she asked as a young founder and asks now of every AI system she touches: not "is it impressive?" but "can we trust it, and can we check?"
Contact Info:
Name: Wenke Du
Email: Send Email
Organization: Wenke Du
Website: https://scholar.google.com/citations?user=3bUZCSEAAAAJ&hl=zh-CN&oi=ao
Release ID: 89203104
Google
RSS