Deep dive · module 90 of 100
AI in production: reliability
Ensure that a AI application remains reliable when it is actually being used.
After this lesson
You can name three points of attention for a AI application in production.
In view · An improvement loop
Reliability remains work
Check normal cases, errors, and similar input.
Watch for deviations, loss of quality, and alerts.
Stop, recover, or use an alternative if quality is insufficient.
Part 1 of 3 · Test before use
Your preference for less movement is on. Choose each part yourself.
Error handling
What happens if the AI service is unavailable or returns an error?
Consistency
Does the system provide similar output for similar input?
Monitoring
How do you know if the quality is declining in practice?
New practice example
From explanation to a usable result.
A reliable application must also react predictably when input, model or connection fails.
Fictional practice situationFictitious: an internal assistant is temporarily unable to retrieve documents. The model itself is accessible.
View a possible solution
This is how it can be done
Have the application report that the source check is currently not possible. Provide a safe follow-up step, such as trying again later or opening the manual itself, instead of producing an unverified answer.
Why this works
Accessibility of the model is not the same as usability of the full service. Make errors visible and recoverable.
Make it slightly more difficult
Write three error scenarios and note per scenario what the user sees and which action is halted.
Extra practice, at your own pace. You do not need to submit anything. This does not count towards an automatic assessment.
Your workplace · learning by doing
Make something. See what changes.
Checking the AI connection… You can already view the written example.
Present your assignment to Ende.
Read the answer, change one thing, and compare.
Previously held
New version
You remain in control
What do you want to improve in your next attempt?
On this device only. AI answers remain only in this page memory. This is self-review, not a grade.
Why this approach works
Accessibility of the model is not the same as usability of the full service. Make errors visible and recoverable.
A step further
Write three error scenarios and note per scenario what the user sees and which action is halted.
Try & discover · no obligation
Control an agent with a control port
Try to send before the check. Then also test a missing source or a malfunction.
- Retrieve source
- Check draft
- Execute action
What is happening?
- No source has been retrieved yet. Feel free to try the submit button first.
All tools, queues and submissions have been simulated. No message is sent and no external service is called.
This trial works locally with fictitious data and does not count towards your lesson completion. It starts anew with a new lesson or reload.
A prompt to get you started
Name three things to watch out for when moving a AI application from experiment to production.
Apply it
A AI service works in a demonstration, but sometimes responds too slowly for the production environment.
Which approach best fits this situation?
- Test load, timeouts, error messages, and a working alternative route.
- Only publish based on the successful demonstration.
- Let users wait indefinitely without notification.
With JavaScript, you can check your approach and receive immediate feedback.
Test your insight
What is an important point of attention for AI in production?
- What happens if the AI service is unavailable.
- Only the speed.
- Nothing, it works automatically.
Extra practice
Choose an application and note one risk to reliability and how you would address it.
Risk: the AI service is sometimes slow. Approach: a timeout and a fallback to a human queue.
