LYNWAY’s Zheng Zhilei on where AI-driven lamp development stands today

LYNWAY's digital twin tools for the ISD interactive display (left) and the intelligent projection headlamp (right). Display effects are compared before a prototype is built.

 

If evaluating a lamp's display once required three to four months of prototyping and validation, could that evaluation begin at the design stage? Zheng Zhilei of LYNWAY showed how his team compares light in a digital twin before anything is built, and how it is connecting AI agents to in-house knowledge and development tools. Engineers can look at more options early in development, but whether the result is accurate and safe on a real road is still theirs to verify.

 

By Sang Min Han _ han@autoelectronics.co.kr
한글로보기

An engineer designs the display image for an automotive lamp. The position of the light sources and the pixel array are fixed, the optical module and software are developed, and a prototype is built. But if, once it is mounted on the vehicle, the text turns out blurry or the graphic breaks apart at the expected distance, the design has to be reworked and a new prototype built.
“It took roughly three to four months, and the cost could exceed 300,000 yuan.”
That is how Zheng Zhilei, head of LYNWAY’s Guangzhou lamp R&D center, described the conventional development process.
He asked what it would mean to see, at the front end of development, the light that could previously be checked only after a prototype existed. The answer he showed in Guangzhou in August 2026 was a pair of digital twins: one for an ISD interactive display, one for an intelligent projection headlamp.
Automotive lamps no longer stop at lighting the road. By controlling individual pixels, they display text, graphics and warning patterns on the outside of the vehicle. The same image can look quite different depending on pixel size and spacing, projection distance, the optical module, brightness and the surrounding environment. Whether an image on the design screen will look as intended on a real vehicle is a separate question.
LYNWAY’s tools let engineers vary conditions such as the display image, viewing distance, screen size, pixel count and size, brightness and color, and compare the results. In the ISD simulator shown on screen, the same content appeared side by side in a color preview and a grayscale preview, with display settings and controls for the optical effect of the lamp cover listed below. Engineers can compare, on a single screen, how the display changes under different settings. On the projection headlamp side, the tool loads projection patterns such as a pedestrian crossing, a construction zone, a snowflake, or a light carpet laid on the road in front of the car, and places the relationship between the optical module and the projection surface in a virtual space. Engineers can screen out design options before committing to a physical prototype.
The three to four months and 300,000 yuan are not figures the digital twin has been proven to eliminate. They are the time and cost of the conventional prototyping and validation process as Zheng described it. Zheng also did not specify which parts of the digital twin tools rely on generative AI and which on conventional simulation. The significance of the approach lies less in a savings rate than in changing the order of validation: moving the moment when results are first seen from after the prototype is finished to the design stage.



Zheng Zhilei said enterprise use of AI is moving from personal prompt engineering to "Harness Engineering," which connects models to knowledge, tools and rules.



From AI That Answers Questions to AI That Uses Tools

Zheng’s scope does not end with digital twins. His question is how to move generative AI from a personal Q&A tool for individual engineers into the company’s development work itself.
Real lamp development involves in-house design data, regulations, test specifications, specialized software and equipment, and review and approval procedures, all working together. Typing good questions into a chat window is not enough to handle that environment.
“We call this kind of enterprise-grade engineering environment ‘Harness Engineering.’”
The harness here is not a vehicle wiring harness. If the model is responsible for judgment, the harness is the foundation that connects the model to the company’s knowledge, tools, rules and workflows. Zheng added one more point: once agents can call tools and correct their own errors, companies need the harness to define the agents’ limits, their tools and their governance.
LYNWAY’s AI build-out began with GPU servers in the third quarter of 2025. After an in-house AI platform and a library of specialized tools, APT Coder, APT Node and an enterprise WeCom bot arrived in the first half of 2026, and the company named an enterprise knowledge base as its third-quarter goal. Zheng described the system as including locally deployed large models, in-house agent tools and an enterprise knowledge base portal.
Zheng introduced APT Node as a “Harness Agent,” in other words a virtual employee. It handles knowledge base Q&A, document processing and the development of automation tools, taking over part of the repetitive daily workload. APT Coder helps engineers analyze projects and write code from requirements stated in natural language, while the WeCom bot answers work questions from all staff and polishes the wording of notices and emails.
He demonstrated screens for generating test reports, analyzing luminance and gamma, and managing simulation jobs. The approach starts by automating repetitive work: gathering data from different instruments, putting it into a common format and turning it into reports.
Zheng said LYNWAY’s agents are entering a third stage, in which agents are genuinely put to work: a closed loop in which the agent plans, executes, checks the results and corrects them on its own. What LYNWAY has today is a set of task-specific tools; connecting them into a single agent system is the next step Zheng envisions.



Right on Screen, Right on the Road?

It is worth separating the judgments a digital twin can move to the front end from the ones that must still be confirmed on physical hardware. On a virtual screen, pixel arrays, the size of a projected image and its expected appearance can be compared quickly. But actual optical performance, the influence of ambient light, distortion caused by the road surface, and human legibility and glare have to go through measurement and in-vehicle evaluation.
In lamp development, then, AI works less as a designer that replaces the engineer than as a way to ask expensive questions earlier. Before choosing which option to prototype, engineers can look at more alternatives and catch obvious problems first. The governance Zheng assigned to the harness has to take concrete form at this stage as well. Deploying the models in-house can be read as a choice to reduce the risk of sending design data to external models. Even so, a record should remain of what data the agent referred to, what it processed automatically, and who checked and approved the result.
The change Zheng described is not limited to the lamps themselves. As more products control light through pixels and software, the tools used to design and validate that light are changing too. Display results that used to be seen for the first time on a prototype three to four months later can now be previewed at the design stage. Whether the light chosen on screen is accurate and safe on a real road is still something engineers have to test and decide.


 

---

About LYNWAY
LYNWAY is an automotive lighting company founded in 2018 by APT Electronics and a Geely-affiliated company, which invested 49% and 51% respectively. The Geely-affiliated stake was later transferred to Yaoning Technology. APT raised its holding to 51% through a capital increase in 2021 and, in December 2023, acquired Yaoning’s remaining 49%, making LYNWAY a wholly owned subsidiary.
APT began at the Hong Kong University of Science and Technology in 2003, established its company in Nansha, Guangzhou, in 2006, and listed on the Hong Kong Stock Exchange in 2024. LYNWAY’s Ningbo plant has a planned capacity of 1.2 million lamp sets a year, and the company says it supplies lamps for Geely, Lynk & Co, Zeekr and smart models. In 2024, APT’s revenue from Geely-related companies was 902 million yuan, 34.8% of the total. APT is also building a production and R&D base in Nansha, Guangzhou, with an investment of about 2.48 billion yuan; planned annual capacity once operational is 700,000 headlamp sets, 700,000 rear lamp sets and 200,000 automotive HUD units.

AEM(오토모티브일렉트로닉스매거진)

---