Katkat
Back

Design Feed Ranking and HTML Generation

FigmaMachine Learning Engineer · Onsite
MediumMachine Learning EngineerOnsiteML System Design

Last updated

This problem tests the design of a dual-purpose ML system: a personalized content feed ranker and an LLM-based component that translates user intent into rendered HTML widgets.

System Design: Feed Ranking and Intent-to-HTML Generation

You are tasked with designing a system for a creative platform where users view a feed of community-generated designs and can use a prompt-based interface to generate custom UI components (HTML/CSS) to embed in their work.

Part 1: Personalized Feed Ranking

Design a ranking system that orders content for a user's feed. Your design should account for:

  • Candidate Generation: How do you retrieve relevant designs at scale (millions of items)?
  • Ranking: Describe the ranking architecture (e.g., multi-stage models, objective functions).
  • Cold Start: How do you handle new creators or users with no interaction history?
  • Evaluation: Define offline metrics (e.g., NDCG, MRR) and online metrics (e.g., Click-Through Rate, dwell time, long-term retention).

Part 2: Intent-to-HTML Generation

Design a service that takes a natural language prompt (e.g., "Create a glassmorphism login button") and generates a functional HTML/CSS snippet.

  • Model Architecture: Choose an appropriate approach (e.g., fine-tuned LLM, RAG for pattern matching, or multi-modal encoder-decoder).
  • Safety & Constraints: How do you prevent the generation of malicious code or non-performant UI layouts?
  • Evaluation: How do you measure the quality and functional correctness of the generated code?

Requirements

  • Provide a high-level system architecture diagram (in text) identifying components like the feature store, model serving layer, and data logging pipeline.
  • Discuss trade-offs between latency and model complexity for both the ranker and the generator.
  • Explain how you would manage data loops: how do you incorporate user feedback (likes, edits, or re-prompts) back into the system?

Hints

Hint 1

For the ranker, consider using a two-tower architecture for fast retrieval, followed by a heavy cross-encoder model for fine-grained ranking.

Hint 2

For the HTML generator, consider using a hybrid approach: a specialized LLM for code generation coupled with a deterministic sandbox environment for real-time functional validation.

Hint 3

For system feedback loops, consider how 'reject' or 'edit' actions in the UI can serve as negative training signals for the code generation model.

Follow-up questions

  • How would you handle 'concept drift' if users' design preferences shift significantly over time?
  • If the HTML generation is too slow, what caching strategies or distillation techniques would you implement to improve latency?
  • How do you design a 'Human-in-the-loop' system to verify generated HTML before it is fully committed to the production design assets library?