Mark Sui

Scope. Reviewed and verified more than 8,000 structured data fields across multiple research domains, checking that each value was supported by reliable evidence and conformed to the expected schema and formatting conventions.

Challenge. Records often combined incomplete context, conflicting claims, and inconsistent naming across English- and Chinese-language sources, making exact verification more important than accepting a plausible-looking result.

How I approached it. I traced individual fields back to authoritative or institutionally credible sources, compared the evidence against the exact submitted value, documented discrepancies clearly, and added stronger supporting links when they were available.

• Audited 8,000+ structured fields across varied research areas, applying consistent review criteria at field level rather than validating records only as a whole.
• Verified names, dates, locations, postal addresses, institutional histories, and classification labels against their cited sources and underlying evidence.
• Flagged entries with missing, weak, outdated, or conflicting support and supplemented them with more authoritative sources when available.
• Maintained standardized formatting and concise review notes so every decision remained traceable and straightforward for another reviewer to reproduce.
• Conducted bilingual web research using English- and Chinese-language government, university, institutional, and cultural-heritage sources.

Scope. Selected for the MOVE AI program to support safety and reliability research for large language models, with work centered on adversarial prompting, evaluation design, and structured review of failure behavior in high-risk scenarios.

Bottleneck. The hard part was that many failures did not appear on short, isolated prompts. Models often looked reasonable at first glance but became inconsistent over longer chains of reasoning, under adversarial reframing, or when the prompt mixed benign context with unsafe intent.

How I approached it. I organized prompts into scenario families, varied wording and escalation patterns, and evaluated not just final answers but whether the model preserved context, grounded facts correctly, and maintained appropriate mitigation behavior as the conversation unfolded.

• Designed adversarial prompts and jailbreak strategies to stress-test agentic model behavior in biosafety and information-hazard scenarios, with emphasis on repeatable rather than one-off failures.
• Conducted systematic QA in Project Castor, reviewing contextual accuracy, factual grounding, refusal behavior, and mitigation quality for harmful or high-risk completions.
• Compared failure patterns across prompt variations to isolate where the model's reasoning diverged, where policy behavior weakened, and which scenarios were most likely to produce unstable responses.
• Worked with a cross-disciplinary team to translate observations into concrete vulnerability reports and recommendations for safer model behavior.

Scope. Contributed to a radar-driven IoT monitoring system designed for elderly health and activity sensing without cameras, helping connect sensing hardware, validation workflows, and real-environment testing.

Bottleneck. The core challenge was that radar signals were sensitive to distance, body orientation, room geometry, and reflective clutter. Respiration and fine motion were especially easy to lose once the environment became less controlled.

How I approached it. I treated the system like a sensing experiment rather than a one-time demo: set up repeatable test conditions, compared logs across placements and environments, and traced instability back to specific physical conditions instead of treating missed detections as random noise.

• Integrated mmWave radar modules to detect movement patterns, restroom usage, and respiration-related signals in privacy-sensitive environments where cameras were not acceptable.
• Calibrated sensing performance under realistic household conditions, testing how results changed with mounting position, target distance, posture, and environmental interference.
• Analyzed logged outputs to identify stability limits, weak-signal cases, and false-detection conditions, then fed those observations back into system tuning and reporting.
• Produced weekly technical summaries that clarified the trade-offs among detection quality, deployment practicality, hardware cost, and privacy requirements.

Scope. Helped build a WeChat-based social product that combined messaging, secure payment, and community features, with work spanning front-end flow, API integration, matching logic, and basic performance tuning.

Bottleneck. The main product issues were friction in the payment flow, match quality that felt inconsistent to users, and response time slowdowns once more ranking logic was layered into the experience.

How I approached it. I treated those as connected product-engineering problems: tighten the interaction path, make backend behavior more reliable, and tune the matching logic using measurable feedback instead of intuition alone.

• Supported front-end interaction redesign and implemented secure payment API integration so the core user flow from browsing to transaction felt more direct and stable.
• Optimized friend-matching similarity weights against profile attributes and observed behavior, improving match accuracy by about 20%.
• Ran performance tests to identify slow points in the request path, then tuned the implementation to reduce response latency by roughly 15%.
• Worked across interface and backend concerns instead of treating them separately, which helped align user experience improvements with system-level changes.