← Meta Interview InsightsSoftware Engineer

Meta·Software Engineer·Onsite - System Design / Architecture·Senior

Senior
Jun 2026

Summary

ML system design round at Meta for a research scientist role. The prompt was about detecting weapon sales, which sounds straightforward until you actually try to spec it out.

Questions Asked (1)

Q1

Design an ML system to detect the sale of weapons on the platform.

System DesignTechnical Trade-offsProduct Strategy
Author's notes

This one had a lot of surface area.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Frame the problem end-to-end by first clarifying scope and success metrics, then design a multi-layered ML pipeline that combines content signals (text, images, video) with behavioral and graph-based signals. Emphasize the trade-off between precision and recall given the high-stakes nature of policy enforcement, and discuss how human review fits into the loop.

Pro tip: Interviewers at Meta want to see that you think beyond model accuracy — proactively discuss adversarial sellers who obfuscate listings (e.g., coded language, cropped images) and how your system evolves to counter evasion, which signals real-world product maturity.

1. Clarify Scope & Define Success Metrics

Confirm what 'weapons' covers (firearms, accessories, illegal modifications vs. legal hunting gear) and which surfaces are in scope (Marketplace, Groups, Messenger). Define metrics such as precision, recall, false positive rate, and policy-violation prevalence to anchor design decisions.

2. Data Collection & Labeling Strategy

Identify training data sources: previously actioned listings, human-reviewed samples, and synthetic adversarial examples. Discuss labeling taxonomy, inter-annotator agreement, and handling class imbalance given that violating listings are rare.

3. Feature Engineering & Multi-Modal Modeling

Design a multi-modal architecture that fuses text embeddings (listing title/description), image/video classifiers (object detection for weapon parts), seller behavioral features (account age, prior violations), and social graph signals (connections to flagged accounts). Use an ensemble or late-fusion approach to combine modality scores.

4. Scoring, Thresholds & Human-in-the-Loop

Output a risk score and route listings into tiers: auto-remove (high confidence), human review queue (medium confidence), and allow (low confidence). Discuss threshold tuning to balance enforcement strictness against false positives that harm legitimate sellers.

5. Monitoring, Feedback Loops & Adversarial Robustness

Instrument the system with dashboards tracking precision/recall over time, appeal rates, and evasion pattern drift. Establish a continuous retraining pipeline fed by reviewer decisions and red-team findings to stay ahead of adversarial sellers.

Key Points to Mention

  • Multi-modal signal fusion: combining NLP, computer vision (object detection), and behavioral/graph features for higher accuracy than any single modality
  • Precision vs. recall trade-off: erring toward higher recall (catching more violations) with human review as a safety net to control false positives
  • Adversarial robustness: handling obfuscation tactics like coded slang, partial images, or private group listings that evade simple keyword filters
  • Cascaded architecture: using cheap, fast models (keyword filters, hashing) as a first pass before expensive deep-learning models to manage latency and cost at Meta's scale
  • Feedback loop and active learning: incorporating human reviewer labels and user appeals back into model retraining to reduce label drift
  • Policy nuance and legal context: distinguishing legal firearm sales (allowed in some jurisdictions) from illegal modifications or unlicensed sales, requiring geo-aware policy logic

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