[책] Generative AI System Design Interview

English study notes on designing GenAI systems: requirements, ML framing, data, models, evaluation, serving and interview questions

GitHub에서 수정

Ali Aminian, Hao Sheng.

The book walks through ten GenAI products, from email autocomplete and translation to chatbots, RAG, image and video generation, using one interview framework: clarify the requirements, frame the problem as an ML task, prepare data, choose and train a model, evaluate it, and design the serving system.

These are my own study notes in English, not a copy of the book. Each note keeps the design reasoning that matters in an interview, an original diagram of the system, and the questions I would expect to be asked, with short answer outlines. Read the book for the full explanations and figures.

• 01. GenAI Fundamentals and the Interview Framework : A deep study companion for the GenAI basics behind every design interview: model families, capacity and compute, scaling laws, data, distributed training, sampling, evaluation, and a seven-step answer framework with a 45-minute plan.
• 02. Gmail Smart Compose : Designing inline email autocomplete in depth: requirements that shape the design, a small decoder-only Transformer, subword tokenization, two-stage training, beam search inside a 100 ms budget, evaluation, and the trigger, generate, filter and rewrite serving path.
• 03. Google Translate : A translation service built on encoder-decoder Transformers: requirements, cross-lingual BPE, denoising pretraining, bilingual versus multilingual finetuning, beam search, BLEU and its successors, serving estimates and interview questions.
• 04. A ChatGPT-Style Personal Assistant : Deep-dive notes on designing a general-purpose chatbot: web data cleaning, decoder-only Transformers with RoPE, pretraining, SFT, reward modeling, PPO and DPO, sampling, layered evaluation, and a guarded, cost-aware serving stack
• 05. Image Captioning for Asset Names : Deep-dive notes on an image captioner that suggests short asset file names: ViT encoder plus pretrained text decoder, CLIP-based data filtering, staged training, beam search with a confidence gate, CIDEr and CLIPScore, and a sized serving path
• 06. Retrieval-Augmented Generation over Company Documents : Deep-dive notes on designing a ChatPDF-style assistant for internal documents: why RAG, layout-aware parsing and chunking, dense and hybrid retrieval, ANN indexes, rerankers, prompt design, RAFT, the RAG evaluation triad, serving, sizing and interview practice
• 07. Realistic Face Generation with GANs : A deep dive on designing an unconditional 1024px face generator: why a StyleGAN-style GAN fits a sub-second budget, data curation and fairness, adversarial training math and stabilizers, truncation and latent editing, FID and its blind spots, and a serving stack with cost estimates
• 08. High-Resolution Image Synthesis with Visual Tokens : A deep-dive on generating 1024px and 2048px scenery with a VQGAN tokenizer and a decoder-only Transformer: token budgets, tokenizer losses, sampling knobs, metrics, and a three-stage serving path with super-resolution
• 09. Text-to-Image Generation with Diffusion : A detailed study of a text-conditioned diffusion system: caption repair, latent U-Net and DiT denoisers, noise-prediction training, guidance and fast samplers, alignment metrics, GPU sizing, and a safety-wrapped serving pipeline
• 10. Personalized Headshot Generation : Turning 10 to 20 selfies into professional headshots: per-user DreamBooth finetuning of a latent diffusion model, prior preservation, identity-gated sampling, face-embedding evaluation, and the batch system around it
• 11. Text-to-Video Generation : Extending a text-to-image diffusion model to five-second 720p clips: video VAE compression, spacetime DiT, joint image-video training, cascades, FVD and CLIP metrics, and a queue-based serving stack with sizing math