Mapping SCALED to RESHADED
Map RESHADED to SCALED: keep your System Design habits, but add model choice, retrieval, token math, and quality evaluation.
We'll cover the following...
You get a design prompt for a customer-support chatbot. You estimate DAU, sketch API servers, add a search index, and call an LLM from one service. The answer sounds organized, but the interviewer asks three follow-up questions you did not plan for. How many tokens does each answer cost at peak load? What hallucination rate is acceptable when the bot answers from policy documents? What happens when the model produces unsafe or uncited advice?
That gap is exactly where many strong System Design candidates stumble. Their classic interview habits are useful, but they need a translation layer before they fit GenAI problems.
This lesson provides that layer by mapping the familiar RESHADED scaffold to the GenAI-focused SCALED framework. The goal is not to replace what you already know. The goal is to preserve your requirement gathering, estimation, and architecture habits, then retarget them toward model behavior, retrieval, safety, and evaluation.
We will not reteach both frameworks from scratch. We will connect them so you can answer GenAI design questions with the same rigor you already use in classic distributed systems.
Recap of the SCALED framework
As discussed previously, SCALED is a six-step framework for designing generative AI systems. It keeps the interview structured, but it shifts attention toward model behavior and output quality.
The six steps are listed below.
System requirements: The system starts by clarifying product behavior, latency, cost, safety, and quality targets.
Choose an AI model: The design selects a model based on context length, reasoning ability, hosting constraints, and failure tolerance.
Acquire and prepare data: The pipeline collects documents, cleans them, chunks them, embeds them, and stores them for retrieval.
Leverage the AI model: The runtime builds prompts, retrieves context, calls the model, applies tools, and handles fallbacks.
Estimate resources: The design computes token volume, serving throughput, storage growth, and hardware needs.
Design the system and evaluate against requirements: The architecture is assembled, then checked against latency, cost, factuality, and safety goals.
In classic distributed systems, the center of gravity often sits around request routing, storage, scaling tiers, and interfaces. In GenAI systems, the request still flows through those layers, but the hard part often shifts to model choice, grounding data, cost per token, and
Note: SCALED extends standard architecture thinking. It does not discard it.
The next sections map these six steps back to RESHADED so your existing interview rhythm still works.
The visual below shows that alignment before we walk through it in detail.
