AI Alumni App, Beijing Powersoft Technology

Chinese universities manage alumni at massive scale — often 100k–200k per school — but lack infrastructure to keep records current, consistent, or useful. As schools pushed to activate alumni networks for fundraising, career services, and community-building, fragmented spreadsheets and WeChat groups weren't enough.

• —Problem statement: Alumni offices had no reliable system for structured alumni data. Core fields like major, industry, and career history were manually entered, inconsistently formatted, and impossible to search or segment at scale.
• —Who the user is: Two distinct users. The primary customer was the university alumni office — institutional teams needing to manage and activate large bases. End users were alumni themselves, especially recent graduates looking to discover classmates' career moves and connect beyond their own cohort.
• —Context & constraints: B2B SaaS delivered via WeChat Mini-App and web for Chinese universities. The biggest constraints were scale, messy legacy data, and source inconsistency — schools often let alumni self-report fields like major, producing hundreds of variants for the same department. I had eight weeks.
• —Why it matters: Without trustworthy data, schools couldn't segment their communities, surface relevant connections, or build useful search. Staff spent time on manual cleanup instead of relationship-building. Alumni missed opportunities for mentorship and community beyond their graduating year.

I combined customer interviews, competitive benchmarking, and workflow analysis to understand both institutional pain points and the end-user experience. Because this was B2B, I focused equally on alumni-facing features and operational burdens: data maintenance, tagging, and trust in the system.

• —User research: ~7 stakeholder interviews with alumni office staff and alumni users. I also analyzed existing data structures and benchmarked against Chinese CRM incumbents, Western alumni tools, and informal WeChat + spreadsheet workflows.
• —Key insight 1: The core customer pain wasn't 'engagement' — it was data quality. Schools wanted a single system that could clean, standardize, and enrich alumni records at scale. Engagement features were only useful if the underlying data was trustworthy.
• —Key insight 2: Alumni users wanted speed, not filter menus. They didn't want to navigate rigid fields to find people — they wanted to describe who they were looking for and get there directly. Natural-language search was the right front door.
• —Key insight 3: Even foundational fields were broken. Letting users self-report majors created hundreds of abbreviations and informal names for the same department — quietly breaking search, reporting, and downstream tagging.
• —Validation: Competitive analysis confirmed a real gap. Local CRM tools had stronger workflows but weaker AI; Western tools had polished UX but no WeChat distribution or Chinese news coverage. The problem was also directly visible in school workflows: questionnaires and manual entry for critical alumni data.

Going to such a prestigious school but not being able to utilize the network beyond those in your year was disappointing.

Eight weeks, fast-moving. I had to balance high-level product decisions with detailed technical tuning — especially in areas where product quality depended directly on AI performance.

• —MVP / prototype: Designed and improved Figma wireframes for the WeChat Mini-App, focusing on a localized, lightweight interface aligned with WeChat usage patterns. Three core flows: news discovery, profile browsing, AI-assisted search.
• —Iteration 1: Early on it became clear that engagement features alone weren't enough — the product needed stronger data quality infrastructure to be useful to schools at all. That shifted my focus toward major normalization and high-confidence matching as foundational product work, not just backend details.
• —What didn't work: Rule-based matching didn't hold up against real school data at scale. User-entered major names without normalization made filtering, reporting, and search unreliable — the kind of quiet failure that breaks trust without obvious error messages.
• —Key challenge: Staying connected to user value while working inside algorithm tuning. Changing one parameter in the normalization pipeline could improve one cluster of matches while lowering performance elsewhere. I had to continuously step back and ask whether the system was becoming more useful for schools — not just more optimized in isolation.