Meta·Machine Learning Engineer·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say

May 2026

Summary

System design round at Meta for an MLE role, focused entirely on building a weapon detection pipeline for ads. Dense question with a lot of moving parts, and I definitely underestimated how much they'd push on the policy and feedback loop side of things.

Questions Asked (1)

Design an end-to-end system to detect advertisements that contain weapons, covering data ingestion, feature extraction across image, text, and video modalities, labeling strategy, model architecture, inference modes, threshold tuning, and how you'd handle new weapon categories over time.

System DesignTechnical Trade-offsAdaptability & Ambiguity

Author's notes

This is a big question and I tried to walk through it linearly which was probably the wrong call.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer as a modular ML system design: start with data ingestion and labeling, then detail multimodal feature extraction and model architecture, followed by inference and threshold tuning, and finally address adaptability to new weapon categories. Emphasize trade-offs at each stage, such as precision vs. recall, latency vs. accuracy, and manual vs. automated labeling. Conclude with a monitoring and feedback loop for continuous improvement.

Pro tip: Demonstrate awareness of Meta's scale and adversarial nature: propose a human-in-the-loop system with active learning to efficiently handle new weapon types, and discuss how to balance false positives (which harm advertiser experience) against false negatives (which create safety risks).

1. Data Ingestion and Labeling

Design a pipeline to ingest ad content (images, text, video) at scale, and propose a labeling strategy that combines automated pre-labeling (e.g., using existing detectors) with human review for ambiguous cases. Consider sampling strategies to handle class imbalance and adversarial evasion.

2. Multimodal Feature Extraction

Extract features from each modality: for images/video, use CNN or vision transformer backbones; for text, use transformer-based embeddings; for video, consider temporal models (e.g., 3D CNNs or frame-level aggregation). Fuse features via early, late, or hybrid fusion.

3. Model Architecture and Training

Choose an architecture that handles multimodality, such as a two-tower model with cross-attention or a multimodal transformer. Train with a multi-task objective (e.g., weapon presence, category, severity) and use techniques like hard negative mining and data augmentation to improve robustness.

4. Inference Modes and Threshold Tuning

Support both real-time (streaming) and batch inference. For real-time, optimize for low latency with model quantization or distillation; for batch, allow more complex models. Tune thresholds per weapon category and modality using precision-recall curves, considering business costs of false positives/negatives.

5. Adapting to New Weapon Categories

Implement a continuous learning loop: monitor for distribution shifts and new weapon types via anomaly detection or user reports. Use few-shot learning or active learning to quickly incorporate new categories with minimal labeled data, and periodically retrain the model.

Key Points to Mention

Multimodal fusion techniques (early, late, hybrid) and their trade-offs
Handling class imbalance and adversarial evasion (e.g., weapons hidden in images or text)
Labeling efficiency: active learning, weak supervision, and human-in-the-loop
Threshold tuning based on business metrics (precision/recall trade-off, cost-sensitive)
Inference optimization for scale: model quantization, distillation, and caching
Continuous learning and adaptation to new weapon categories via few-shot or meta-learning

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.