AI Development (Internship) — GenAI Assistants for Community Services

Two lines of work, one product mindset: • Interview Intake & Triage - Built a structured intake for client interviews: eligibility, urgency, domain, and privacy dimensions. - Operationalized scoring rubrics and routing rules so requests auto-triage to the right assistant or human queue. - Turned interview insights into personas, intents, and slot models to drive conversation design. • Assistant Design & Delivery - Designed multi-turn flows (greeting → discovery → task → confirmation → handoff) with clear fallbacks and escalation. - Authored reusable prompt templates and tool-use protocols; added retrieval hooks and small knowledge indices. - Provided bilingual (EN/ZH) phrasing guidance and a terminology/glossary layer to improve clarity for seniors and non-native speakers. - Unblocked technical issues across the team: RAG recall gaps, latency spikes, hallucination hotspots, and role/permission config. - Instrumented logs and lightweight analytics (success, deflection, fallback rate, escalation) and wrote runbooks for ops. - Delivered demo scripts and admin training so non-technical staff can configure scenarios and permissions safely.

• Product over model: start from user goals and constraints; flows, tone, and fallbacks make the experience—not model selection. • Measure adoption, not just accuracy: task success, first-response resolution, fallback/escation rate, and time-to-value guided iterations. • Guardrails win trust: role-based access, retrieval boundaries, refusal styles, and human-handoff rules reduced operational risk. • Language and accessibility matter: bilingual prompts, simpler sentence patterns, and a shared glossary improved comprehension. • Delivery loop: interviews → flow/prototype → pilot → analytics/QA → enablement. This kept stakeholders aligned and the bot useful.