[3/3] System Design. Preparing for a complex GenAI interview (Category SystemDesign)

Finishing the book (1 and 2)It's worth seeing. 10 tasks prepared by the authors for training

1️⃣ Gmail Smart Compose A system that offers a continuation of the phrase right at the time of writing. Here, it’s important to keep up with very little delay, show a hint only when the model is confident enough, and further filter out failed, toxic, or inappropriate options. 2️⃣ Google Translate A machine translation system that takes text in one language and turns it into text in another language. The main questions are how to work with different languages, how to learn from multilingual data, and how to measure the quality of translation when literal translation is not always the best. 3️⃣ ChatGPT-like Personal Assistant A personal AI assistant who conducts a dialogue, remembers the context, can access external tools and adapt to the user. Security, memory management, privacy and control over what the assistant can do on behalf of the person are particularly important. 4️⃣ Image Captioning A system that looks at an image and generates a text description. This is an example of a multimodal problem: the input is a picture, the output is text. It is important not just to “guess objects”, but to describe the scene in a way that is useful to the user. 5️⃣ Retrieval-Augmented Generation A system that answers questions not only from the “memory” of the model, but first searches for relevant snippets in documents, a knowledge base, or corporate search. The main topics are: how to break documents into parts, how to search for fragments that are close in meaning, how to make the model rely on the sources found and how to show links to them. 6️⃣ Realistic Face Generation A system for generating realistic faces. It is important not only the quality of the image, but also risks: biases in data, the generation of unwanted content, the possibility of abuse and the need for protective restrictions. 7️⃣ High-Resolution Image Synthesis A system for generating or improving high-resolution images. The main engineering complexity is that such operations are expensive for calculations, so you often have to build a multi-step pipeline: first rough generation, then improvement, detailing and increasing resolution. 8️⃣ Text-to-Image Generation A system that creates an image from a text description. Here, it’s important to understand how a text query turns into a visual result, how to manage image style and detail, and how to filter out prohibited or insecure queries and results. 9️⃣ Personalized Headshot Generation A system that generates a personalized user portrait, such as a business avatar or profile photo. The main difficulties: to maintain recognition of a person, not to violate privacy, properly store and delete user images, as well as to prevent abuse of someone else's identity. 🔟 Text-to-Video Generation A system that creates video from a text description. This is one of the most difficult classes of tasks: you need not only to generate beautiful shots, but also to maintain the coherence of the scene in time, the movement of objects, style, characters and at the same time manage very expensive and long calculations.

These tasks should not just be read, but solved as training interviews - set a timer and sketch the design yourself: requirements, ML-formulation, data, model, metrics, architecture, deployment and monitoring. And then compare with the author's analysis.

#SystemDesign #AI #GenAI #Architecture #Engineering #ML #Interview #Software