CrackML by @ml.with.umang
Interview questions / ML System Design
ML System Design interview question

Design a content moderation ML system

Design an end-to-end machine learning system for content moderation. Clarify the policy target, labels, model architecture, online/offline evaluation, serving, human review, feedback loops, abuse adaptation, and monitoring.

hardsystem designEvidence 77/1001 source reportGoogle

The 60-second answer

Define the moderation policy first: taxonomy, severity, latency, human-review/appeal path, and false-positive vs false-negative costs. Build policy-consistent labels with hard negatives; use staged filtering/classification with calibrated confidence and human review for uncertain or high-impact cases.

Build the answer in this order

1
Frame the problem

Define the moderation policy first: taxonomy, severity, latency, human-review/appeal path, and false-positive vs false-negative costs.

2
Design the data path

Build policy-consistent labels with hard negatives; use staged filtering/classification with calibrated confidence and human review for uncertain or high-impact cases.

3
Choose the modeling stack

Evaluate class-wise precision/recall, calibration, policy-weighted cost, and important safety slices—not a single aggregate accuracy number.

4
Serve, evaluate, iterate

Monitor prevalence, reviewer disagreement, appeal reversals, drift, latency, adversarial adaptation, and feedback contamination.

A useful interview mental model

This is the shape of a strong answer—not a script to memorize.

01Requirements
02Data
03Model / Retrieval
04Serving
05Monitor

Senior-level signal

  • Separate policy correctness from model accuracy and version them independently.
  • Design for adversarial users, delayed ground truth, reviewer quality, and safe rollback when enforcement rates shift.

What the interviewer is really testing

Product framing, data design, modeling choices, serving constraints, reliability, evaluation, and explicit trade-offs.

Likely follow-up questions

What changes at 10× traffic or data volume?
Which failure mode would you monitor first in production?
How would you evaluate this offline and online before rollout?

Common weak-answer patterns

  • Jumping to a model before defining the product contract.
  • Listing components without bottlenecks, metrics, or failure handling.
  • Ignoring data quality, serving latency, monitoring, and iteration.