Katkat
Back

Design an image-to-multilanguage translator

AxonSoftware Engineer · Technical Screen
MediumSoftware EngineerTechnical ScreenSystem Design

Last updated

This question tests the ability to design a distributed, scalable image processing pipeline that integrates OCR with translation services while addressing latency, storage, and reliability concerns.

Problem Statement

Design a service that allows users to upload an image containing text (e.g., a menu or a sign) and receive the text translated into their preferred language. The system must support thousands of concurrent requests, provide low-latency responses, and handle images in various formats (JPEG, PNG, HEIC) up to 10MB each.

Functional Requirements

  • Users can upload an image and specify a target language.
  • The system returns the extracted text and the translated version.
  • The system supports a set of 50+ target languages.

Non-Functional Requirements

  • Latency: The initial response for text extraction should be fast; translation can be processed as a secondary step.
  • Reliability: Handle service failures in third-party OCR or translation APIs.
  • Scalability: The system must scale based on traffic spikes (e.g., travel season).
  • Privacy: Ensure user images are handled securely and deleted or anonymized according to policy.

Deliverables

  1. High-level architecture diagram (or component description).
  2. API design (endpoints and request/response payloads).
  3. Data model for storing user requests and metadata.
  4. Strategy for handling long-running processing tasks and asynchronous communication.
  5. Trade-offs between choosing an off-the-shelf cloud OCR API vs. a self-hosted ML model.

Hints

Hint 1

Consider using an asynchronous task queue or message broker to decouple image upload from the heavy processing logic.

Hint 2

Think about how to provide an immediate user experience (e.g., returning a task ID and polling/webhooks) vs. a synchronous response.

Hint 3

Address how to cache results for common translations or identical images to save on expensive API costs or compute power.

Follow-up questions

  • How would you handle images with multiple languages present, or text that is handwritten rather than printed?
  • If the translation service has a strict rate limit, how would you design your system to prioritize requests and manage backpressure?
  • How would you monitor the system to detect if the OCR model's accuracy starts degrading over time?