Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Alumnium - Open-source AI-powered Test Automation

Alumnium - Open-source AI-powered Test Automation

A talk about Alumnium, an open-source AI-powered test automation tool, and its comparison to other tools in the space.

This talk was given at the AI Camp meetup on February 13, 2025 (https://www.aicamp.ai/event/eventdetails/W2025021317).

Recording - https://youtu.be/O8YTgCeXiA8?t=2273.

Alex Rodionov

February 13, 2025
Tweet

More Decks by Alex Rodionov

Other Decks in Technology

Transcript

  1. ALEX RODIONOV - Quality Architect @ Toptal - Author @

    Maccy - Ruby maintainer @ Selenium - Ruby maintainer @ Bazel
  2. 1. Writing tests: a. fi guring out the interactions and

    locators b. writing the code 2. Supporting tests: a. adapting to UI changes b. dealing with fl akiness c. surviving massive refactoring PAIN OF END-TO-END TESTING?
  3. 1. Autocomplete and chat (Copilot et al). 2. AI-powered test

    generators (LaVague). 3. AI-powered test runners (Shortest, Alumnium). AI-POWERED SOLUTIONS
  4. 1. De fi ne tests in Cucumber format. 2. lavague-qa

    generates Selenium tests using AI. 3. Commit the tests to your repo. 4. Run on CI. 5. Re-generate whenever something changes. LAVAGUE (TEST GENERATOR) https://www.lavague.ai/
  5. + You can see what’s actually generated + Once generated,

    the tests are fast + Cheap - Cannot be integrated into existing tests - Have to use Gherkin language - No auto-regeneration (i.e. if the test failed) LAVAGUE (TEST GENERATOR) https://www.lavague.ai/
  6. 1. De fi ne tests in Shortest format. 2. Commit

    them to repo. 3. Run on CI. 4. Test uses Anthropic Computer Use API to interact with browser. SHORTEST (TEST RUNNER) https://shortest.com/
  7. + Easy to use + Adapts to changes in UI

    - Cannot be integrated into existing tests - Expensive - Slow https://shortest.com/ SHORTEST (TEST RUNNER)
  8. 1. High-level API for Selenium/Playwright: a. do something on the

    page; b. get something from the page; c. check something on the page. 2. Integrates into existing tests. 3. Provides gradual migration path. 4. Supports all major AI providers (Anthropic, Google, OpenAI). 5. Works on low-tier models (gpt-4o-mini, gemini-2.0- fl ash, claude-3-haiku) 6. Works with open models too (Llama 3.2). 7. Open source under MIT license. ALUMNIUM https://alumnium.ai/
  9. PERFORMANCE (1 TEST) 1. LaVague = 51s to generate, 1s

    to run 2. Shortest = 58s to run 3. Alumnium = 8s to run https://alumnium.ai/
  10. COSTS (1 TEST) 1. LaVague = $0.08 (pay once) 2.

    Shortest = $0.15 3. Alumnium = $0.001 (80-150 times cheaper) https://alumnium.ai/
  11. 1. LLM lack of determinism. 2. Support for di ff

    erent models. 3. Prompting . 4. Scaling to bigger a11y trees. 5. Improving performance. 6. Working with sites that has poor a11y. CHALLENGES https://alumnium.ai/
  12. 1. Other programming languages (JavaScript, Ruby, etc.) 2. Mobile applications

    testing. 3. Local models. 4. Evaluation benchmarks. ROADMAP https://alumnium.ai/