Kevin Zhou

Led end-to-end design, training, and deployment of a voice-enabled tutoring assistant integrating Whisper ASR, TTS, and conversational agents, supporting bilingual instruction across English and Mandarin. Built a multi-agent tutoring orchestration framework using LangChain + LlamaIndex to dynamically adapt content, covering ~12,000 knowledge points. Constructed and maintained a robust RAG pipeline using FAISS / vector databases to retrieve curriculum content, reducing hallucination rates by ~40%. Managed distributed training and experiment workflows via Ray + MLflow, reducing iteration cycles from ~10 days down to 3 days. Collaborated with curriculum, pedagogy, and assessment teams to align AI feedback with standardized test formats, improving pilot student pass rates by ~15%. Mentored junior engineers, established reproducible experiment logging, and instituted CI/CD pipelines for model rollout.

Led the design and implementation of AI-assisted CRM/support modules, integrating conversational agents + CRM workflows to resolve ~70–80% of inbound customer issues autonomously. Built context-aware response pipelines by combining user metadata, entity grounding, and fine-tuned LLMs, reducing escalations by ~18%. Integrated AI modules into Kustomer's CRM platform for context-sensitive replies. Developed tooling for non-technical support operators to adjust conversational flows and rules, doubling the throughput of updates. Performed error analysis and iterative model refinement, improving intent/response accuracy by ~20%. Instrumented internal analytics to track KPIs (deflection rate, fallback rate, average time to resolution).

Designed and implemented recommendation engines (user-to-item, session-based, vector embedding retrieval models) for web and mobile clients, boosting click-through/engagement metrics by 8–12%. Built offline and online inference pipelines with caching and latency optimizations to support real-time recommendation. Developed feature engineering frameworks (behavioral signals, time-decay, recency) and model evaluation pipelines, increasing model refresh velocity by ~30%. Integrated A/B testing platforms to validate recommendation strategies in production with thousands of users. Monitored model drift and retraining triggers using MLOps tooling, ensuring model freshness and relevance over time.

Developed internal caregiver scheduling, billing, and payroll software (Python + Django + MySQL), reducing administrative work by ~25%. Automated form routing, notifications, and email workflows (Zapier/cron), saving staff ~8 hours/week. Created dashboards to visualize performance metrics: caregiver punctuality, client satisfaction, completion rates. Designed secure role-based access control for staff and clients, ensuring data privacy compliance. Facilitated data migrations and managed incremental feature rollout to avoid service disruption.