SWShuoxin Wangintrtc.hashnode.dev·Aug 14 · 11 min readDon’t Open Another AI Inbox: Build a Question Queue on Top of Community ChatYour users are already explaining what they need in group chat. The tempting response is to add an AI bot, a feedback inbox, or a dedicated support room. That creates a second problem: someone must co00
SWShuoxin Wangintrtc.hashnode.dev·Aug 13 · 13 min readStream the Preview, Commit the Message: A Reliable Architecture for AI-Assisted ChatA streaming assistant creates an uncomfortable engineering tension: users expect words to appear immediately, but a social chat system must preserve coherent history, moderation decisions, consent, re00
SWShuoxin Wangintrtc.hashnode.dev·Aug 11 · 12 min readDon’t A/B Test the LLM: Randomize the Voice Conversation PolicyA new LLM can sound clearly better in a demo and still make a voice companion worse in production. The tension is not simply model quality versus model cost. In a real-time conversation, changing the 00
SWShuoxin Wangintrtc.hashnode.dev·Aug 5 · 11 min readBeauty AR Needs a Degradation Receipt, Not Just an FPS CounterA Beauty AR demo can look perfect on the device used to build it. Production asks a less comfortable question: what happens when makeup, segmentation, stickers, and video processing compete for a smal00
SWShuoxin Wangintrtc.hashnode.dev·Jul 27 · 12 min readThe Model Is Not the Control Plane: Designing AI-Assisted Feedback TriageYou want AI to reduce the time spent sorting and answering user feedback, but you do not want a model quietly deciding which messages matter—or sending a polished, incorrect answer to a real person. T00
