Approach. I designed an AI workspace that brings together wallet-domain documentation, query tools and troubleshooting SOPs. A RAG layer grounds answers; the dialogue flow routes harder cases to humans.

Outcome. Routine consultation and diagnosis became self-service, creating a reusable “knowledge → answer → human fallback” operating loop.

Intelligent order search

Designing for the trade users remember—not the code they forget.

Problem. Analysis of 200+ support tickets showed users remembered the underlying company but not the derivative ticker. Exact-code search left them dependent on customer support.

Approach. I defined a three-layer retrieval system: clearer filters, Chinese/English fuzzy semantic matching and underlying-to-derivative association.

Outcome. Search success rose 27%, average lookup time fell from 12 to 7 seconds, daily PV increased from 24K to 28K and related complaints dropped to zero.

Content distribution engine

Making AI content production measurable end to end.

Problem. Manual distribution was slow, content quality varied and user acquisition could not be reliably attributed.

Approach. I connected five modules—tasks, AI-assisted production, channel distribution, monitoring and settlement—and introduced base plus performance incentives.

Outcome. One campaign generated 3,288 pieces, 12.64M impressions and 8,449 effective search referrals; trial acquisition reached 10K new users per week.