← all talks

Deploy Anything with AI: From Elixir to Rust to Edge Functions

Deploy Anything with AI: From Elixir to Rust to Edge Functions. By Erwan Rougeux. October 28, 2025, Iglu Bangkok, Thailand. Fork It Community.

Who am I?

VP, International Growth @ Clever Cloud (sovereign PaaS).

Passionate about software development & cloud architecture; open-source and privacy advocate.

Love privacy-by-design & new AI tools.

Cartoon avatar of Erwan

De-rusting with AI

  • Context: I wanted to “de-rust” my coding muscles without actually coding.
  • Constraint: ChatGPT 3.5 only; I copy-pasted outputs, no deep refactors.
  • Challenge: a niche domain (handpan tablature) + a web UI, in 3 sleepless nights.
ASCII handpan notes and chords chart ChatGPT generating a handpan tablature for the Super Mario Bros. theme

The rules & setup

Ground rules

  • Tooling: ChatGPT 3.5 (no plugins, no external copilots) + plain browser / dev server.
  • Method: rapid prompt → copy → run; fix only what breaks.
  • Scope: minimal web app that renders handpan tabs and lets me poke at the UX.
  • Risk accepted: messy code, fragility, zero tests.

Why this matters

  • Forces AI to design + implement in a niche.
  • Surfaces what LLMs do well vs. where human engineering is still required.

What v1 (AI-made) actually delivered

The MVP

  • A simple browser app that can generate & display handpan tabs.
  • Basic interactions, good enough for a live demo.

Strengths

  • Speed: from idea to demo in hours, not weeks.
  • Coverage: boilerplate UI, glue code, small utilities on demand.

Limits

  • Brittle: hidden coupling, inconsistent naming, unclear data model.
  • Maintainability: no tests, no architecture, hard to extend.

MVP

The v1 app: a page of handpan diagrams showing which notes to strike

What I learned

Where AI shines

  • Scaffolding, boilerplate, CSS/DOM snippets, quick polyfills.
  • Explaining unknown APIs well enough to unblock me.

Where AI struggles

  • Robust domain modeling (musical edge cases, layouts).
  • State management & testing discipline.
  • Non-functional needs: perf, security, accessibility.

Working strategy

  • Small, testable prompts; lock the spec; iterate with examples.
  • Treat the model like a junior: review, constrain, verify.

The upgrade path

The v2.0

  • From toy to tool with Windsurf + Claude Code.
  • Architecture: typed core, clear modules, DB, tests.
  • UX: snappier rendering, cleaner editor, export / share flows.
v2: tablature library, light theme v2: tablature library, dark theme v2: MuseScore import screen

Questions?