When choosing my undergraduate research direction, the field was wide open and no one was going to tell me which choice was "correct." I narrowed it through three layers of filtering. Trend — AI agents and embodied intelligence are durable directions, not fad waves, so I targeted robotic security. Infrastructure — the foundation had to be an open-source, hardware-decoupled ecosystem with long-term staying power, the way Linux became for operating systems; that pointed me to ROS 2. Whitespace — papers on agent-driven offensive security were exploding, but defensive work was nearly empty. I use LLMs daily and know they have "jagged intelligence" — strong in some dimensions, weak in others — meaning the attacker agent itself has exploitable weaknesses. A USENIX Security paper using Unicode honeytokens to counter LLM agents confirmed this direction was academically viable. These three layers converged into ROS2Pot — the first DDS-native, physics-aware honeypot for ROS 2 targeting LLM-agent attackers.

In the capstone project for the Agentic AI & Workflow course, we had to find a Malaysian local business, diagnose its operational bottlenecks, and deploy AI into its actual workflows using n8n. I led an internationally diverse team. Rather than splitting work evenly, I started by mapping each person's edge: some had local business connections, some had strong documentation skills, some had sharp technical execution — and I placed each at their strongest position while I held the overall direction and filled any gap. We delivered a B2B+B2C sales enablement platform, and every member of the team got an A.

Prof. Xu Rongsheng — "father of China's Internet," China's first anti-hacker, Senior Researcher and Doctoral Supervisor at CAS-IHEP, Chief Scientist at the IHEP Network Security Lab, and recipient of the National Science & Technology Progress Award (Special Class) — visited my university for a lecture. The minutes immediately after his talk were my only window to get direct feedback from someone of his standing. I stepped up and used 90 seconds to present three sets of facts: First, the trend — papers on agent-driven offensive attacks were exploding. Second, the insight — "jagged intelligence" means attacker agents themselves have exploitable weaknesses, so countering agents with a honeypot is technically viable. Third, the stakes — robotics and embodied intelligence are reaching real deployment, making this a live attack surface, not a toy problem. After listening, he gave me this evaluation: "Yucheng thinks innovatively; against agent attackers in robotic cybersecurity, he has proposed an interesting approach."

I see myself as a generalist who can step in and take over wherever needed — if any part of the work stalls, I can pick it up and finish it. This shapes how I view collaboration: good teamwork isn't dividing work evenly, it's putting each person at their strongest position so the team actually delivers. The international team in the Agentic AI & Workflow course is a good example. Instead of equal splits, I first mapped each member's edge — the one with local business contacts led BD, the strong writer owned requirements docs and final deliverables, the strongest engineer built the n8n pipelines. I kept directional authority and operated as the team's floating role — wherever something broke, I patched it. We didn't ship a deck; we shipped a working B2B+B2C sales enablement platform. Every member got an A — no one was averaged out by the team.

Recomby.ai's AI Employee architecture. Most of the GEO (Generative Engine Optimization) market splits into two camps: pure SaaS tools (which leave all the work to the customer to figure out) or traditional agencies (which absorb all the work into human labor). I rejected both based on one core judgment: SEO optimizes keyword matching and can be fully automated; GEO optimizes the semantic representation of the business itself, which only works when "GEO experts" and "business experts" collaborate. So I designed the division of labor into two clean layers: we, as the GEO domain experts, only build skills that empower the agent — pipelines, domain prompts, professional tools; the customer, as the expert on their own business, just hands over the facts of their business to the system. The product form is a Feishu CLI (carrying data and scaffolding) + a Codex-class agent — the agent is what actually does the work. The result is a crushingly favorable cost structure for the customer: compared to hiring a full-time GEO employee, our offering is subscription-level overhead; compared to traditional agencies charging monthly for human labor, an agent doing the work brings the price down another tier. The customer doesn't need to become a GEO expert, doesn't need to maintain an in-house GEO team, and doesn't pay agency-level human markup — they just pay for skills + tokens.

ROS2Pot is a large topic — protocol layer, physics layer, agent counteraction, attacker classification, ATT&CK mapping. Polishing every dimension to perfection before submitting to a top-tier venue would take 1.5 to 2 years and risk being scooped by peers. I made a deliberate trade-off: package the proven core — protocol-layer indistinguishability under professional ROS 2 scanning tools, physical consistency between the simulator and the real system, and a counteraction experiment that successfully elicited the attacker agent's operational objective — into a 3-page CCS poster submission first, and defer the rest to future work. This wasn't a compromise forced by limited capacity; it was a deliberately designed two-stage roadmap: claim the academic foothold during undergrad, then compound on it through grad school and PhD.

ROS2Pot required a full stack I didn't originally have: ROS 2, DDS-RTPS, rclpy, Fast-DDS, and PyBullet physics simulation. My starting point was just Linux and Python basics. My strategy was learn by doing — for each new component I'd spin up a minimum runnable demo first, then dig into documentation only when I hit a concrete problem; far faster than reading documentation linearly. The hardest layer to crack was DDS-RTPS — the communication middleware protocol underneath ROS 2. The applied result is a high-interaction, physically consistent honeypot: protocol-layer indistinguishability under professional ROS 2 scanning tools, and physical consistency between the simulator and the real system.

In my freshman year, I weighed about 135 kg (around 297 lbs). At the start of my sophomore year, I set myself an aggressive target: lose 30 kg in two years — a level no one around me had pulled off. Instead of crash dieting, I built a sustainable system: black coffee in the morning to anchor caloric intake, two light meals a day to maintain a steady deficit, resistance training to preserve muscle, and regular badminton with my roommate to turn cardio into a social commitment I couldn't skip. Two years later I stabilized at 105 kg — no meal replacements, no medication, just compounding returns on diet structure and training discipline.