Giles Lavelle

• Led the ground-up rebuild of retool.com from a legacy Gatsby/JS stack to TypeScript, Next.js, and Tailwind, reaching 99.999%+ uptime and 95–100 Lighthouse scores (was 20–30)
• Built the web platform’s CI and CD pipelines from scratch as sole engineer: 1,000+ unit tests and Playwright checks on every PR, and automated deploys across GitHub Actions, Vercel, Slack, and Linear
• Led adoption of Claude Code in the web codebase and published @lavelle/lint-agent, a linter for agent skills now used across Retool’s engineering org

Projects:Site searchApp gallery

Multimodal conversation practice for language learners, by voice and text: an AI partner you talk to out loud, and a separate AI teacher whose notes sit beside the conversation. Five language pairs, 27 partner characters.

• Built Offbook, a multimodal AI language tutor, solo in 7 weeks: ~77,000 lines of TypeScript (19,000 of tests), 295 commits, 74 PRs, deployed on Cloudflare Workers
• Designed a 10-role multi-agent LLM pipeline with persistent per-character memory, proactive push messaging, and adaptive difficulty, across Claude, Gemini, and self-hosted models
• Built a voice and text pipeline with streaming AI responses (live transcription, STT and TTS across 7 and 9 providers, including local Whisper and Qwen3-TTS) and an MCP server with OAuth
• Built a prompt evaluation framework: 27 promptfoo suites, 215 cases, an LLM-as-judge, and per-suite cost tracking, used to compare models and refine every prompt

Projects:Offbook NumbersOffbook TranslateOffbook