My year-long journey living in China’s bustling capital. From cutting-edge technology and work life to cultural discoveries and other adventures, here is where I will share some of the experiences, challenges, and insights gained while living at the crossroads of innovation and tradition.

Settling Down

It was a hectic start to the year. I had just returned to Singapore after spending my winter holiday in Boston with my Dad, still shaking off the jet lag while juggling visa paperwork, packing up, and making time to see family and friends before leaving for a year.

I arrived in Beijing in early January 2025, stepping into winter's grip with temperatures hovering around freezing. The city felt both familiar and foreign—I had grown up in Shanghai, but Beijing was a different beast entirely. The dry cold, the sprawling hutongs, the political weight that seemed to hang in the air.

Finding an apartment, setting up a bank account, getting a local SIM card, completing health checkups and registering for a residential permit—the logistics of settling into a new city consumed the first few weeks. But there was also the excitement of starting fresh, of being in a place where AI was moving at breakneck speed and opportunities felt tangible.

Getting Started

Work began in mid-February at Nolibox's office in 东升大厦. The first week was a whirlwind of onboarding, meetings, and absorbing as much context as I could about the company's products and technical stack.

Beyond work, I threw myself into Beijing's ecosystem. I enrolled in evening classes at Peking University (PKU) outside of the NOC program, attended seminars and conferences whenever I could, and tried to say yes to every invitation that came my way. The density of opportunities in Beijing was remarkable—on any given day, there were multiple events worth attending, from intimate founder dinners to large-scale tech conferences.

Something about student life, I joined PKU's 百团大战 and was impressed by the vibrancy of student life. Many societies and clubs were professionally organized, some with industry experts on their advisory boards.

The PKU Golf Society trial session introduced me to local/international master's and PhD students as well as exchange students. Interestingly, almost everyone I spoke to was curious about my role as an AI engineer, and nearly all were considering pivoting to something AI-related.

The Beijing Rhythm

Life in Beijing settled into a rhythm. Wake up, commute to the office (sometimes by subway, sometimes by shared bike when the weather permitted and sometimes by DiDi), work on technical projects, attend events or classes, read or catch up on the latest AI developments.

The food was different from Shanghai—more oily and heavier. I found my favorite breakfast spots, my go-to lunch food deliveries on 美团, and my preferred coffee shops.

What struck me most was the pace. Beijing's tech scene moved fast—maybe not as polished as Silicon Valley, but with a raw energy and willingness to ship that felt distinctly Chinese. "先做出来再说" (build it first, talk later) was the prevailing ethos.

Work Experience

I accepted the offer to work for Nolibox (北京计算美学科技有限公司), a company founded by Xu Zuobiao (CEO) and Huang Shengyu (CDO) in 2020. The founders came from Tsinghua and Peking University, and two of them were recognized on the 2022 Forbes 30 Under 30 Asia list in Enterprise Technology. The company had secured nearly millions in RMB in funding and worked with several Fortune 500 companies and State-Owned Enterprises.

Phase 1: Building Core Tools

The GPT-Powered Image Selection Tool

My first major project was to enhance the productivity of the designers. For context, we had thousands of images being generated each day by AI models. Imagine a platform similar to Canva, these are AI-generated posters, images, and templates featured on landing pages. Each day, designers had to verify image quality, ensure there were no sensitive content, and pick the best-looking images.

I started by understanding how designers perceived quality from the angle of contrast, composition, legibility. I noticed that many images had spelling errors and grammar mistakes that the native Chinese designers couldn't spot in English text.

At that time, the best multi-modal model was GPT-4o. I built a tool that leveraged GPT's vision capabilities to automatically analyze and select images. The challenge was translating subjective concepts like "aesthetically pleasing" into something more concrete. For example, I scoped it to "aesthetically pleasing for marketing purposes" and carefully structured the prompts to output JSON that could integrate directly with the company's Lark-based workflow.

Because these tasks are high in volume but not necessarily required to be real-time. I eventually proposed using OpenAI's Batch API which reduced costs by 50%. When GPT-4o started exhibiting "sycophantic" behavior, I monitored the situation and migrated to GPT-4.1.

Template Parser and Office Move

The second major project was building a JSON template parser to bridge PosterMyWall and Polotno Studio for MolyPix.AI, Nolibox's Western-facing SaaS product. Polotno's documentation was insufficient, many optional elements were undocumented, so I spent considerable time debugging through trial and error. The end result was a single-click tool that designers could use directly from Lark tables.

During this period, the company moved from 东升大厦 to 中关村东升科技园二期. Colleagues and I packed our desks into boxes on Friday before the move. The new office was in a well-built tech park, surprisingly crowded at lunch despite the baren surroundings.

YOLO Object Detection for 1688.com

I also worked on improving object detection for a critical 1688.com project. The existing YOLO model had been trained on fewer than 300 images. I analyzed which product categories and layouts underperformed, gathered and labeled images to build a dataset of over 1,000 images, and recategorized them into a proper 70-20-10 training/testing/validation split.

The improved model achieved an mAP@50 of 75.5%, Precision of 77.5%, and Recall of 67.5% with labels for border, design element, text, and watermark-compared to the original single "design element" label.

This resulted in much more accurate detection of the words, labels, design elements, main product/figures/objects on the project listing images. As such, we could do a much better post processing and regenerate a better product listing.

Email Marketing and Other Initiatives

Beyond technical work, for the overseas product, I noticed the underutilization of user data. There was no email campaign for new user onboarding, and the existing welcome email needed improvement. I proposed this to the CEO, who agreed and gave me the greenlight to spearhead R&D for email marketing using Mailjet and precise customer outreach using Apollo.

One example was using Apollo/Clay/Dripify to target marketing managers, and I customized personalized email templates/messages using prompt engineering and LLMs.

Looking for new ways to generated editable AI designs, I went on to research on the capabilities of existing models on generating HTML based designs. Back then it was Claude 3.7 Sonnet that was one of the best in class, it does a fairly decent job at generating HTML/CSS based posters, but I added additional MCPs (Model Context Protocols) to it, like heroicons-mcp and lucide-icons-mcp for icon/logo and it was even better.

Phase 2: AI Agent Development

By mid-2025, the industry's shift toward agentic applications was obvious, a trend amplified by releases like Lovart.AI, the first AI Design Agent. The founder wanted to try something new, a social media content generation agent. I led the development of the agentic system, a content generation system with integrated memory and recommendation services.

We worked on using RAG (scraping and vectorising historical data of user's social medias to understand how they post, the tone, the length and what kind of design aesthetic they enjoy) and Mem0 for more personalised content. Websearch tools are powered by the likes of Tavily, Crawl4AI and Firecrawl.

The challenges were significant. Maintaining context accuracy while expanding memory features required careful redesign of update logic. Reference image vectorization for personalization proved complex, as achieving accurate template matching needed extensive tuning. For example, how should the user uploaded image be used in accordance with the template image that we searched for at the back.

Introducing Agile Methodology

With the sole Product Manager leaving in early July, I noticed the critical gap in user research and product direction. I took the initiative to introduce Agile methodology to the team, acting as Scrum Master for weekly sprint ceremonies. This improved engineering alignment and task visibility significantly.

However, I also learned the limits of frameworks. During one retrospective, higher management mentioned that it "sort of" covered everything. But Sprint Retrospective and Sprint Review are fundamentally different, retrospectives are for team members to share what went well, went wrong, or could improve. In a small team where founders are also product owners, these distinctions get blurred. Theory doesn't always translate smoothly to practice.

Other Technical Deep Dives

• LangFuse SDK integration for prompt management and agent tracing
• DeepEval implementation for LLM evaluation using G-Eval with evaluation steps
• n8n workflows for automated SEO blog generation and a competitive intelligence monitoring system
• Sanity CMS setup for the website blog and Generative Engine Optimisation (GEO)
• Video transcript tools for the agent, supporting YouTube, TikTok, Facebook, Instagram, and X/Twitter.

Travelling

One of the greatest privileges of being based in China for a year was the access to travel. High-speed rail made weekend trips effortless, and I made it a point to explore as much as I could beyond Beijing.

Shanghai

Shanghai held a special place in my heart—I had spent my childhood there from 2007 to 2013, walking through 南京路步行街 to get to school in 黄浦区. Returning as an adult, I was struck by how much had changed.

My first trip back was during the 清明 (Qingming) holiday in April, combined with company visits to Meituan, NIO, and Baidu Apollo. I also visited my old primary school and was amazed to see AI already integrated into the curriculum. This reinforced my belief that China could leverage AI to its fullest by enabling its people with the latest technology from a young age.

I visited 外滩 (the Bund) at night, watching the light show dance across Pudong's skyline. The contrast with my childhood memories was stark—the city felt more confident, more global, more technologically ambitious.

One trip took me to 临港新片区 (Lingang Special Area), a special economic zone focused on institutional innovation. Tesla's megafactory and the Chinese-made aircraft factory are situated there. The goal is 800,000 long-term residents by end of 2025, scaling to 1.5 million by 2035. The location is strategic—close to the port and Pudong Airport—but amenities are sparse compared to the city center. Still, I enjoyed the peacefulness and noted down tech companies there for future reference.

Hangzhou

I visited Hangzhou in late July for the AdventureX hackathon. The city has a different energy from Beijing or Shanghai—greener, more relaxed, yet still deeply tech-forward.

AdventureX 2025 was China's largest ever hackathon with 5 days of event and dozens of workshops on tech, VC, entrepreneurship, and innovation.

Being the only software engineer on my team, I built a presentable demo and pitched during the expo. We created "Agentic Factoria," an AI system for industrial factory management powered by SUPCON's open-source supOS platform with built-in MQTT for real-time data handling. The system used LLM-driven agents to tackle factory management, fault fixes, and logistics.

I have also been to a handful of other places like Nanjing, Suzhou, Heibei (Yuxian 蔚县), Chengdu, Chongqing, Guilin, Yang Shuo.

Silicon Valley: A Contrasting Perspective

During China's National Day holiday in October, I made a deliberate choice to spend two weeks in San Francisco, to meet friends in SF and attend the SF Tech Week rather than traveling for pure leisure.

I stayed in Menlo Park with a group of friends from NOC Silicon Valley. The house had 7 NOC SV students, all driven, with clear goals, mostly thinking about their own startup ideas. I also followed them to their entrepreneurship classes at Stanford.

At SF Tech Week (organized by a16z), sessions like "Agentic AI: What's Hot & What's Not" reshaped how I evaluate AI startups. The consensus was clear: successful agentic AI products don't just demonstrate intelligence-they reliably get real work done within existing systems.

The fireside chat with Bolt.new founder Eric Simons emphasized "speed with discipline"-building quickly without losing sight of customer reality. Concepts like "community as distribution" and user-led marketing showed me why many technically strong products fail to scale.

OpenRouter CEO Alex Atallah shared real-world data from processing over 50 trillion tokens-operational trade-offs around cost control, latency, tool calling reliability, and model orchestration. This wasn't abstract discussion; it was production reality.

Company Visits, Conferences, Seminars and Hackathons

Beyond the day-to-day work, I made a conscious effort to immerse myself in China's broader tech and entrepreneurial ecosystem. The density of events in Beijing was remarkable—there was always something worth attending.

Seminars and Talks

马玉山院士: 产研学深度融合创新的实践与思考

One of the most memorable talks was by Academician Ma Yushan (马玉山院士), one of only five Chinese Academy of Sciences academicians from a 民营企业 (private enterprise) background. His perspective on integrating industry, research, and academia for practical innovation was eye-opening. He spoke candidly about the challenges of translating academic research into real-world products—a gap that felt particularly relevant to my work.

王磊: Business Model Innovation Series

I attended two talks by Wang Lei (王磊):

• 创业导论：商业模式创新与股权融资 — He shared multiple use cases of AI agents in China and was refreshingly open about US advances in AI and robotics, providing a balanced global perspective.
• AI时代商业模式创新与企业资本运作密码 — This focused on business model types and equity/stock distribution methods for structuring businesses. Essential knowledge for anyone thinking about starting a company.

2025"MEET未来"系列大课明势：宏观研判

This forum featured Tsinghua University professors sharing their views on the economy. It was shockingly candid—the first speaker even discussed potential errors in the government's GDP calculations. This level of open discourse surprised me and challenged my assumptions about intellectual freedom in China.

「融合·共生」2025 首都高校创业者年会

A conference jointly organized by multiple academic institutes. This was my first time interacting with a humanoid robot in person. The founders who spoke were remarkably open during Q&A, welcoming audience members who expressed interest in collaboration.

梁宁：从混沌到清晰，探寻商业与人生的真需求

Liang Ning's seminar on finding genuine needs in business and life was one of the most thought-provoking. She has a reputation for deep strategic thinking, and her frameworks for understanding user needs and market dynamics stayed with me throughout the year.

2025 GAI+ Conference

A focused gathering on generative AI applications. The conversations here were highly technical, discussing everything from model fine-tuning to deployment challenges to emerging use cases across industries.

Fireside Chat with Economic Counsellor Chua Zong Lun

This session provided insight into the Singapore government's collaborations and planned opportunities for Singaporeans in China and the ASEAN region. Understanding the diplomatic and economic layer added context to my personal career considerations.

VC Guest Speakers

During one of the classes, two guest speakers from the VC scene shared their perspectives. I took detailed notes on the differences in financing options for startups in China versus the US—the emphasis on growth metrics versus profitability, the role of government-backed funds, and the expectations around timelines.

Company Visits

Meituan (美团) — Shanghai

Meituan's scale is staggering—from food delivery to hotel bookings to bike-sharing, they've built an everything-app for services. The visit focused on their data-driven approach to optimizations.

NIO (蔚来) — Shanghai

The electric vehicle company's Shanghai facility showcased their battery swap technology and approach to customer experience.

Baidu Apollo — Shanghai

Baidu's autonomous driving division demonstrated their progress on L4 autonomy. This was my first time riding in a robotaxi through Shanghai's streets.

科大讯飞 (iFlyTek) — Beijing

I had always thought of iFlyTek as merely a hardware company selling AI-enabled devices. I was shocked by the breadth of their LLM applications across industries—education, healthcare, legal, and even collaboration with Shanghai's Nuclear Research Center. Their voice recognition technology in Mandarin is genuinely world-class.

Hero Games (英雄游戏) — Beijing

Eight levels of their Beijing HQ occupied by 1,200+ employees. The visit was organised by our mentor – Shannon Cheung, an impressive angel investor.

Final Reflections

Looking back on 2025, it's hard to distill a year into neat lessons. But a few things stand out:

On China: The country is more complex than any single narrative suggests. It's simultaneously a place of incredible competitive intensity and frustrating bureaucracy, of genuine warmth and, of ancient tradition and breakneck modernization. Anyone who claims to fully understand China is probably wrong.

On travel: Seeing China beyond Beijing was essential. The country is vast and varied—Guilin's karst mountains feel nothing like Shanghai's skyline or Chengdu's teahouses. Each place offered a different lens on Chinese life and culture.

On community: The connections made this year—with colleagues, with fellow NOC students, with founders and investors and academics—will outlast any single project or achievement. Investing in relationships was never wasted time.

On myself: I've become more comfortable with ambiguity, more confident in my ability to figure things out, and more clear on what I want to build with my career. Not all the answers, but better questions.

On cultural adaptation: Building globally competitive products requires more than cutting-edge AI-it demands localized marketing, cultural sensitivity, and user-centric design.

On organizational resilience: Document processes, cross-train team members, and establish sustainable structures early, even when resources are limited.

A year in China, 2025. It was everything I hoped for and nothing like I expected.