
Siyan Li
Nonclinical AI chatbots for pharmacies and clinics
Habilidades

Conheça meus serviços


Portfólio
Experiência profissional
AI Chatbot Developer — Non-Clinical Pharmacy FAQ Assistant
Self-Employed / Freelancer • Autônomo
Nov 2026 - Present • 0 mos
Designed and deployed a working, publicly accessible chatbot demo for a community pharmacy scenario. The bot answers administrative FAQs — opening hours, prescription drop-off, insurance basics, delivery, parking — and refuses all clinical questions, redirecting patients to a pharmacist. The core value is boundary design: the bot is trained to stay strictly on the administrative side, never give medical advice, never check symptoms, and never invent an answer outside its knowledge base. Clinical questions receive a fixed refusal and a redirect to a pharmacist or doctor. Built with a static HTML frontend, Cloudflare Pages and Pages Functions as a CORS-safe proxy, and a Z.AI language model backend with environment-variable key management. The demo is live and used as a reference implementation for pharmacy and clinic front-desk automation.
Independent Developer & Product Owner — China Trend Radar Tracker
Self employed / Freelance • Autônomo
Nov 2026 - Present • 0 mos
Solo-designed, built, and currently maintain a production web platform that tracks whether a Chinese short-video format is genuinely being replicated across independent creators, and whether it is still worth entering. The system collects data from multiple social platforms, runs local relevance and traffic pre-filters to avoid unnecessary AI cost, then uses a vision-language model to suggest match strength. The core design principle is that AI only suggests — every confirmed trend requires a human decision. No trend, cluster, or recommendation is ever auto-confirmed by the model. Built with vanilla HTML/CSS/JavaScript (no framework, no build step), Vercel Serverless Functions, Supabase Postgres with row-level security, and scheduled background workers. Added provider usage logging, daily cost guards, failure-explicit data handling, and an offline test suite covering 600+ cases. Currently in active use.