Generative AI Application Lifecycle

Six stages that repeatCopy Anchor Link

The story: Running a delivery business with a hired driver:

1. Decide what you'll deliver, to whom, and what customers expect.
2. Hire an experienced driver, or raise and train one from scratch if nobody suitable exists. That depends on how unusual your deliveries are and how much you know about your routes.
3. Make them better: clearer instructions, a map book, route training, and helpers who handle multi-stop trips.
4. Test them: ride along, compare them against standard driving tests, and track delivery times automatically.
5. Put them on the road.
6. Collect customer feedback and delivery data, then retrain or replace as needed.

When a new road opens or a better driver becomes available, you go back and repeat steps.

In AI/AWS terms: The lifecycle is iterative: stages get revisited as needs change or better models appear.

1. Define a use case: the problem, requirements, and stakeholder expectations.
2. Select a foundation model: use a pre-trained model or build one from scratch, depending on whether a suitable model exists, how complex the use case is, and how much domain data you have.
3. Improve performance: prompt engineering, RAG, fine-tuning, and agents.
4. Evaluate: human evaluation, benchmark datasets, and automated metrics.
5. Deploy: integrate the model into the target environment.
6. Monitor and improve: collect feedback, usage data, and metrics, then retrain, fine-tune, or update the model.