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

SeniorPrefer not to say

Jun 2026

Summary

Meta MLE onsite focused entirely on a single open-ended ML system design problem around weapon sales detection on a marketplace. Dense question, lots of ground to cover, and I definitely didn't get to all of it cleanly.

Questions Asked (1)

Design an ML system to detect high-frequency weapon sale listings on a marketplace platform. Walk through training data collection and labeling, whether funnel-stage models share training data, feature engineering, model architecture choices, evaluation metrics given the high-stakes nature of the domain, deployment, and feedback loops.

System DesignTechnical Trade-offsProduct Analytics & Metrics

Author's notes

This is a lot to fit into one session.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around the ML lifecycle, emphasizing the high-stakes nature by prioritizing precision and recall trade-offs, and discuss how each component (data, features, models, deployment) addresses the unique challenges of detecting weapon sales. Highlight the importance of human-in-the-loop and feedback mechanisms to continuously improve the system.

Pro tip: Demonstrate maturity by acknowledging the ethical and legal implications, and propose a tiered review system where high-confidence predictions trigger immediate action while lower-confidence ones are queued for human review, balancing automation with safety.

1. Data Collection and Labeling

Gather a diverse dataset of listings, including positive examples of weapon sales and hard negatives (e.g., toy guns, accessories). Use a combination of automated filters and human annotators with domain expertise to label data, ensuring high-quality ground truth.

2. Feature Engineering and Model Architecture

Engineer features from text, images, user behavior, and network signals. Choose models that handle multimodal data, such as a fusion of text and image models, and consider a cascade of models from cheap to expensive for efficiency.

3. Training and Evaluation

Train models with a focus on high recall to catch potential violations, but maintain precision to avoid false positives. Evaluate using metrics like precision-recall AUC, and set thresholds based on business requirements for action.

4. Deployment and Feedback Loops

Deploy in a staged manner, starting with shadow mode, then human review, and finally automated actions. Implement feedback loops where human decisions and user reports are fed back to retrain models, and monitor for drift and adversarial attacks.

Key Points to Mention

Multimodal feature engineering: combining text (listing description), images (weapon detection), and user metadata (history, network).
Model architecture: consider a two-stage approach with a fast filter (e.g., logistic regression on text) followed by a more complex model (e.g., CNN on images) for high-recall candidates.
Evaluation metrics: prioritize recall at high precision (e.g., precision at 99% recall) due to high stakes, and use confusion matrix analysis to understand failure modes.
Deployment: use a tiered action system (e.g., auto-remove, human review, monitor) based on confidence scores, and ensure low latency for real-time detection.
Feedback loops: incorporate human reviewer decisions and user reports as labels for continuous training, and use active learning to prioritize uncertain cases.
Ethical and legal considerations: ensure compliance with laws, avoid bias, and maintain transparency in decision-making.

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