kuanzw

Responsible for the development of AI Tutors and AI Voice Coaches. Leveraged LLMs to analyze online data, create new agents, and integrate image understanding, which increased the AI tutor's response rate from 53.36% to 80.62%. Fine-tuned the response model through a complete pipeline: prompted the generation of real-world Q&A data, used an LLM-as-a-judge for data classification, cleaning, and evaluation, performed Supervised Fine-Tuning (SFT), and deployed the model, leading to improved Q&A accuracy. Developed an AI workflow to automatically analyze over 14,000 characters of text in 3 days (a task that would typically take one week), extracting 226 intents and 799 corresponding Q&A pairs for the chatbot application.

Developed and maintained the game backend distributed system using C++, Python, Go. Designed and implemented a scalable matchmaking system using Redis for low-latency player matching across multiple game modes, ensuring it could handle high volumes of traffic during peak hours (over 3000 qps in brench test). Developed and deployed microservices for leaderboards, player progression, and in-game economy using Docker and Kubernetes. Using Mongo as the database for storing and persisting data related to leaderboards, player progression, and the in-game economy, taking advantage of its document-oriented data model.

Responsible for developing features across different server clusters, utilizing MySQL as the primary database and Redis for caching. Refactored and fixed core server components, following the STL (Standard Template Library) standards to restructure the skip list data structure, improving code quality and maintainability. Optimized performance in multi-player combat scenarios: Conducted load testing with multiple bot players simulating online environments. Used the perf profiling tool to identify hot spots in multi-player combat scenarios. Optimized damage calculation logic, reducing the overhead of a hot spot function by 19%.