--- sheet_format: 1.0 id: 2025-10-forkit-bangkok-derusting-with-ai title: "Deploy Anything with AI: From Elixir to Rust to Edge Functions" speaker: Erwan Rougeux with: [] date: 2025-10-28 event: "Fork It" place: "Iglu Bangkok, Bangkok, Thailand" format: talk language: en duration: "unknown (no recording)" video: null slides: https://slides.wanoo.me/2025-10-forkit-bangkok-derusting-with-ai/ topics: [ai-assisted coding, chatgpt, claude code, windsurf, handpan, mvp, software engineering] source: "Summary written by Claude (AI) from the slides only; there is no recording of this talk. Not a verbatim transcript; what was said on stage may differ." review: spot-checked review_date: 2026-10-17 --- # Deploy Anything with AI: From Elixir to Rust to Edge Functions **In one sentence:** Erwan Rougeux recounts building a handpan tablature web app by copy-pasting ChatGPT 3.5 output, then rebuilding it "from toy to tool" with Windsurf and Claude Code, to show what AI does well and where engineering still matters. **Note: there is no recording of this talk.** This sheet is built from the slide deck alone, so it reflects what is written on the slides, not what was said on stage. The outline uses slide numbers instead of timestamps. ## TL;DR Erwan, VP International Growth at Clever Cloud, wanted to "de-rust" his coding muscles without actually coding. He gave himself strict rules: ChatGPT 3.5 only, no plugins or copilots, copy-paste the output, fix only what breaks, and build a niche app (handpan tablature with a web UI) in three sleepless nights. The v1 MVP worked well enough for a live demo and went from idea to demo in hours, but it was brittle and had no tests, no architecture and an unclear data model. His lessons: AI is good at scaffolding, boilerplate and explaining unfamiliar APIs, but weak at domain modelling, state management, testing and non-functional needs. Treat the model like a junior: review, constrain, verify. The v2.0, rebuilt with Windsurf and Claude Code, has a typed core, clear modules, a database and tests. ## Erwan's key points (From the slides only; there is no recording.) - The experiment was to "de-rust" his coding skills without writing code himself, using only ChatGPT 3.5 and copy-paste, with no deep refactors (slides 3–4). - He picked a niche domain on purpose, handpan tablature, to force the AI to design and implement something it has rarely seen (slides 3–4). - He accepted the risks up front: messy code, fragility, zero tests (slide 4). - The AI-made v1 delivered a simple browser app that generates and displays handpan tabs, from idea to demo "in hours, not weeks" (slide 5). - Its limits: hidden coupling, inconsistent naming, an unclear data model, no tests and no architecture, which made it hard to extend (slide 5). - Where AI shines, per the slides: scaffolding, boilerplate, CSS/DOM snippets, quick polyfills, and explaining unknown APIs well enough to get unblocked (slide 7). - Where it struggles: robust domain modelling (musical edge cases, layouts), state management and testing discipline, and performance, security and accessibility (slide 7). - His working strategy: small, testable prompts, a locked spec, iteration with examples, and treating the model like a junior developer (slide 7). - The v2.0 went "from toy to tool" with Windsurf and Claude Code: typed core, clear modules, a database, tests, faster rendering, a cleaner editor, export/share flows and MuseScore import (slide 9). ## Outline - Slides 1–2: Title and speaker introduction: VP International Growth at Clever Cloud (sovereign PaaS), open-source and privacy advocate who likes new AI tools. - Slides 3–4: The challenge and the rules: ChatGPT 3.5 only, prompt, copy, run, fix only what breaks; a minimal handpan-tab web app in three sleepless nights; messy code accepted. - Slides 5–6: What the AI-made v1 delivered: a working MVP built fast, but brittle and hard to maintain; screenshot of the v1 app. - Slide 7: Lessons learned: where AI shines, where it struggles, and how to work with it (treat it like a junior). - Slides 8–9: The upgrade path: v2.0 rebuilt with Windsurf and Claude Code, with proper architecture, tests and better UX; screenshots of the tablature library and MuseScore import. - Slide 10: Questions. ## Notable quotes Slide text only (there is no recording): > "I wanted to “de-rust” my coding muscles without actually coding." — slide 3 > "Surfaces what LLMs do well vs. where human engineering is still required." — slide 4 > "Treat the model like a junior: review, constrain, verify." — slide 7 > "From toy to tool with Windsurf + Claude Code." — slide 9 ## Questions this answers **Q: What does Erwan Rougeux think about coding with AI?** A: According to his slides, AI is fast and useful for scaffolding, boilerplate and explaining unfamiliar APIs, but it struggles with domain modelling, state management, testing, performance, security and accessibility. He recommends treating the model like a junior: review, constrain, verify (slide 7). **Q: Can you build an app with ChatGPT 3.5 by copy-pasting its output?** A: Erwan's experiment says yes, for an MVP: a browser app that generates and displays handpan tablatures, built in three nights, good enough for a live demo. The result was brittle, with no tests and no architecture (slides 3–5). **Q: What is the handpan tablature app Erwan built?** A: A web app that renders handpan tabs, showing which notes to strike. v1 was made with ChatGPT 3.5; v2.0 was rebuilt with Windsurf and Claude Code and adds a tablature library, light and dark themes, export/share and MuseScore import (slides 6, 9). **Q: How do you turn an AI-generated prototype into a maintainable tool?** A: In Erwan's v2.0, by adding what the AI prototype lacked: a typed core, clear modules, a database and tests, built with Windsurf and Claude Code, plus small testable prompts and a locked spec (slides 7, 9). **Q: Is there a video of "Deploy Anything with AI" at Fork It Bangkok?** A: No. There is no recording of this talk; only the slide deck is available. ## Mentioned Clever Cloud, Fork It, Iglu Bangkok, ChatGPT 3.5, Windsurf, Claude Code, MuseScore, handpan, Super Mario Bros. (example tablature) ## Slides outline 1. Deploy Anything with AI: From Elixir to Rust to Edge Functions — title, Iglu Bangkok, 28 Oct 2025, Fork It 2. Who am I? — VP International Growth at Clever Cloud; privacy-by-design and new AI tools 3. De-rusting with AI — context, ChatGPT 3.5 copy-paste constraint, handpan tablature challenge in 3 nights 4. The rules & setup — tooling, method, scope, accepted risks, and why it matters 5. What v1 (AI-made) actually delivered — MVP, strengths (speed, coverage), limits (brittle, unmaintainable) 6. MVP — screenshot of the v1 handpan diagrams 7. What I learned — where AI shines, where it struggles, working strategy 8. The upgrade path — section divider 9. The v2.0 — from toy to tool with Windsurf + Claude Code; typed core, DB, tests; UX improvements 10. Questions?