San Francisco Bay Area
Recent: voice AI infra, LLM systems research, networking at Arista. Past life in architecture and design 🌐 kevwjin.com | 📩 kevwjin [at] gmail [dot] com
• 0→1 product engineering for voice AI (React Native, Supabase, Fly.io) • Built custom voice conversation system (STT-LLM-TTS pipeline), reducing costs by 10x and improving response quality by 25% compared to OpenAI’s Realtime API • Designed LLM-as-a-judge evaluation harness to stress-tests AI with 30 parallel simulated users
• Built LLM-enabled static analysis system to automatically fix resource leaks in normalized Java programs • Constructed representative dataset (rNJR) of 100 resource leaks for evaluating LLM repair performance, where the LLM-enabled system achieved 82% fix rate, outperforming the state-of-the-art RLFixer tool by 16% • Discovered 2 novel leak patterns (Properties and Runnable leaks) representing 62% of previously unfixable cases by analyzing compilation and fix validation results across multiple static analyzers
• Improved hardware-test execution speed for 100+ engineers by optimizing 40+ router configuration calls • Built parser to extract hardware (ASIC) configuration data from YAML files at compile-time • Accelerated future product development cycles by creating reusable SysDB library from parsed hardware data • Built CLI tool used by 50+ engineers to query the hardware data, replacing manual database queries • Refactored existing test suites for better long-term unit and integration testing support
• Saved configuration time for 8000+ customers by restricting CLI to valid router commands • Reduced network latency by implementing BGP-LU-to-ISIS tag redistribution, enabling faster packet routing between different network segments • Improved tracking of network connectivity and failures by incorporating seamless Bidirectional Forwarding Detection (sBFD) support in interop echo mode • Maintained system modularity for 50+ engineers by dynamically determining the network protocol (IPv4 or IPv6)
Built web-crawler scraping 250+ webpages per hour and storing data in AWS S3 for COVID wearables study