InterviewStack.io LogoInterviewStack.io

Meta Applied Scientist (Mid-Level) Interview Preparation Guide

Applied Scientist
Meta
Mid Level
6 rounds
Updated 6/19/2026

Meta's Applied Scientist interview process for mid-level candidates consists of an initial recruiter screening followed by a technical phone screen and a full-day onsite loop of 4-5 interviews. The process evaluates your ability to conduct applied research, develop novel algorithms, implement solutions at scale, and communicate findings. Each round assesses different competencies: research reasoning, technical implementation, ML systems design, experimental validation, and cultural fit. Meta values candidates who can bridge research and engineering by taking abstract problems and delivering production-ready solutions.

Interview Rounds

1

Recruiter Screening

2

Technical Phone Screen

3

Onsite Round 1: Applied Research and Problem Formulation

4

Onsite Round 2: Machine Learning Systems Design

5

Onsite Round 3: Machine Learning Implementation and Coding

6

Onsite Round 4: Statistical Analysis, Experimentation, and Behavioral

Frequently Asked Applied Scientist Interview Questions

Model Deployment and Inference OptimizationHardTechnical
22 practiced
Evaluate the cost and energy trade-offs of moving inference for parts of a popular mobile app from cloud GPUs to on-device inference across 10 million monthly active users. Describe the metrics and assumptions you would use, outline how to compute a break-even point (cost per request vs device battery impact, development and maintenance costs), and discuss privacy, UX, and regulatory considerations.
Feature Engineering and Feature StoresEasyTechnical
61 practiced
Define feature freshness and staleness. Provide three concrete metrics you would track to measure feature freshness for both online and offline stores. Explain how those metrics inform operational decisions (e.g., eviction, re-materialization frequency).
Machine Learning System ArchitectureMediumSystem Design
19 practiced
Design an end-to-end ML pipeline for a fraud-detection service that ingests 2M events/hour, stores ~10TB raw/day, and must serve 500K online inference requests/sec globally with P50 latency 50ms. Show components for ingestion, storage, feature-store, distributed training, and serving; provide data flow and justify choices for streaming vs batch, consistency, and scalability.
Experimentation Methodology and RigorHardTechnical
82 practiced
Design an experiment to estimate heterogeneous treatment effects (HTE) for a homepage redesign. Describe how you would (a) pre-specify segments vs use data-driven methods, (b) estimate HTE (e.g., causal forests, uplift models), and (c) validate discovered heterogeneity to avoid spurious claims.
ML Algorithm Implementation and Numerical ConsiderationsMediumTechnical
79 practiced
Implement a numerically stable function in Python (NumPy) that computes log-sum-exp across axis=1 for a 2D array of logits. The function must handle very large and very small logits without overflow/underflow. Also describe time and memory complexity.
Model Deployment and Inference OptimizationEasyTechnical
18 practiced
Explain horizontal versus vertical scaling for model serving. Describe autoscaling triggers (CPU/GPU utilization, request latency, queue depth), scaling granularity, and the trade-offs between scaling up (larger machines) and scaling out (more replicas) in terms of cost, cold-start time, and fault tolerance. Give an example autoscaling policy suitable for a latency-sensitive inference service.
Feature Engineering and Feature StoresMediumSystem Design
67 practiced
Design a simple dependency graph representation for feature computations. Describe the data structures you would use to represent nodes (features), edges (dependencies), and metadata such as last materialized timestamp and compute cost. Explain how you would detect cycles and compute an efficient recomputation schedule when a low-level source table changes.
Machine Learning System ArchitectureEasyTechnical
17 practiced
When packaging a trained model for production, what artifacts should you include (model weights, preprocessing code, encoders, feature definitions, metrics, and env spec)? Explain how a model registry and artifact storage support reproducibility, deployment gating, and rollback in an ML lifecycle.
Experimentation Methodology and RigorEasyTechnical
67 practiced
Define statistical power, Type I and Type II errors, significance level (alpha), and minimum detectable effect (MDE) in the context of online A/B testing. Then, given a baseline conversion rate of 5% and a desired two-sided alpha=0.05 and power=80%, explain qualitatively how the required sample size changes as you (a) halve the MDE, (b) change to a one-sided test, and (c) increase desired power to 90%. No calculations required — focus on intuition and trade-offs.
ML Algorithm Implementation and Numerical ConsiderationsEasyTechnical
88 practiced
Write pseudo-code (or Python) for numerically stable softmax and cross-entropy loss computation for a multiclass classifier. Explain why naive implementations can lead to overflow and how the stable version prevents it.

Want to create your own tailored preparation guide using our deep research?

Get Started for Free

Interview-Ready Courses

Visual-first, interactive, structured learning paths

Browse Applied Scientist jobs

AI-enriched listings across hundreds of company career pages

Explore Jobs
Meta Applied Scientist Interview Questions & Prep Guide (Mid-Level) | InterviewStack.io