The decomposition interview is the signature FDE interview format. You're given a vague business problem and must break it into technical components, propose an architecture, and discuss trade-offs — all while communicating clearly.

These problems are organized by industry and difficulty. Each includes a problem statement, key questions to ask, and a suggested approach.

How to Approach Decomposition Problems

The Framework (5 steps, 45 minutes)

1. Clarify (5 min) — Ask questions. Understand the customer, constraints, and success metrics.
2. Decompose (10 min) — Break the problem into 3-5 sub-problems.
3. Prioritize (5 min) — Which sub-problem delivers the most value first?
4. Design (15 min) — Architecture the solution. Draw diagrams. Discuss data flow.
5. Trade-offs (10 min) — What could go wrong? What would you do differently with more time?

Logistics & Supply Chain (Problems 1-10)

Problem 1: Package Routing Optimization

Scenario: A shipping company delivers 500K packages daily across 200 cities. They want to reduce delivery times by 15%.

Key questions to ask:

• What data do they currently collect? (GPS, timestamps, weather, traffic)
• What's their current routing system? (Manual? Basic algorithm?)
• What does "delivery time" mean? (Warehouse to door? Last mile only?)

Approach:

• Sub-problems: (1) Data ingestion from GPS/IoT, (2) Route optimization algorithm, (3) Real-time re-routing, (4) Performance monitoring dashboard
• Start with: Historical data analysis to identify bottleneck routes
• Tech: Graph optimization, real-time streaming (Kafka), geospatial queries (PostGIS)

Problem 2: Warehouse Inventory Prediction

Scenario: An e-commerce company has 15 warehouses. They over-stock 30% of items and under-stock 20%. Design a system to predict optimal inventory levels.

Problem 3: Fleet Maintenance Scheduling

Scenario: A trucking company has 3,000 vehicles. They want to predict maintenance needs to reduce breakdowns by 50%.

Problem 4: Port Container Tracking

Scenario: A shipping port processes 10,000 containers daily. Containers get lost or delayed. Build a real-time tracking system.

Problem 5: Last-Mile Delivery Optimization

Scenario: A grocery delivery service operates in 5 cities. Each driver makes 20-30 deliveries per shift. Optimize driver assignment and routing.

Healthcare (Problems 6-15)

Problem 6: Patient Readmission Prediction

Scenario: A hospital network wants to reduce 30-day readmission rates. They have 5 years of patient records across 12 hospitals.

Key questions to ask:

• What data is available? (EHR, labs, medications, demographics)
• HIPAA constraints? Data residency requirements?
• What interventions are possible if we predict high risk?

Approach:

• Sub-problems: (1) Data pipeline from EHR systems (HL7/FHIR), (2) Feature engineering, (3) ML model (XGBoost or similar), (4) Clinical dashboard for care teams, (5) Feedback loop for model improvement
• Start with: Retrospective analysis on historical readmissions
• Critical: Model explainability — clinicians need to understand WHY a patient is flagged

Problem 7: Medical Image Triage

Scenario: A radiology department processes 500 scans daily. They want AI to prioritize urgent cases.

Problem 8: Drug Interaction Alert System

Scenario: A pharmacy chain wants real-time alerts when prescriptions have dangerous interactions.

Problem 9: Clinical Trial Patient Matching

Scenario: A pharma company has 50 active clinical trials. They need to match eligible patients faster.

Problem 10: Hospital Bed Capacity Planning

Scenario: A 500-bed hospital frequently runs at 95%+ capacity. Design a prediction system for bed availability.

Financial Services (Problems 11-20)

Problem 11: Fraud Detection Pipeline

Scenario: A fintech processes 10M transactions daily. Current fraud detection catches 60% of fraudulent transactions with a 5% false positive rate. Improve both metrics.

Key questions to ask:

• Latency requirements? (Real-time blocking vs. post-transaction review?)
• What data is available? (Transaction details, device info, user behavior, merchant data)
• What happens when fraud is detected? (Block, flag, require verification?)

Approach:

• Sub-problems: (1) Real-time feature computation, (2) ML scoring pipeline, (3) Rules engine for known patterns, (4) Investigation dashboard, (5) Feedback loop from investigators
• Architecture: Streaming pipeline (Kafka → feature store → model serving → decision engine)
• Key trade-off: Latency vs. accuracy. Adding more features improves detection but increases latency.

Problem 12: Know Your Customer (KYC) Automation

Scenario: A bank spends 45 minutes on average per KYC check. They want to automate 80% of checks.

Problem 13: Portfolio Risk Dashboard

Scenario: An investment firm manages $5B across 200 portfolios. Build a real-time risk monitoring system.

Problem 14: Loan Default Prediction

Scenario: A lending platform wants to predict loan defaults at application time.

Problem 15: Anti-Money Laundering (AML) Graph Analysis

Scenario: A bank needs to detect suspicious transaction networks across 50M accounts.

Defense & Government (Problems 16-25)

Problem 16: Satellite Imagery Analysis

Scenario: A defense agency receives 10TB of satellite imagery daily. They need to detect changes (new construction, vehicle movement) automatically.

Problem 17: Cybersecurity Threat Intelligence

Scenario: A government network operations center monitors 500K endpoints. They want to reduce mean time to detect threats from 72 hours to 4 hours.

Problem 18: Disaster Response Resource Allocation

Scenario: After a natural disaster, coordinate rescue teams, supplies, and medical resources across an affected region.

Problem 19: Border Surveillance System

Scenario: Monitor 500 miles of border using a combination of sensors, cameras, and drones.

Problem 20: Supply Chain Security for Critical Infrastructure

Scenario: A government agency needs to verify that hardware components haven't been tampered with across a global supply chain.

Retail & E-Commerce (Problems 21-30)

Problem 21: Real-Time Pricing Engine

Scenario: An e-commerce platform with 5M products wants dynamic pricing based on demand, competition, and inventory.

Problem 22: Customer Segmentation at Scale

Scenario: A retailer with 20M customers wants to create dynamic segments for personalized marketing.

Problem 23: Store Layout Optimization

Scenario: A grocery chain wants to use purchase data and foot traffic to optimize product placement.

Problem 24: Returns Prediction and Prevention

Scenario: An online fashion retailer has a 35% return rate. Predict and reduce returns.

Problem 25: Omnichannel Inventory Visibility

Scenario: A retailer with 500 stores and an e-commerce site wants unified, real-time inventory visibility.

Energy & Manufacturing (Problems 26-35)

Problem 26: Predictive Maintenance for Wind Turbines

Scenario: A wind farm operator has 200 turbines with IoT sensors. Predict failures before they cause downtime.

Problem 27: Energy Grid Load Balancing

Scenario: A utility company needs to balance supply and demand across a grid with 30% renewable (intermittent) energy.

Problem 28: Quality Control with Computer Vision

Scenario: A manufacturing line produces 50,000 units daily. Detect defects using camera inspection.

Problem 29: Digital Twin for Factory Operations

Scenario: Build a digital twin of a factory floor to simulate and optimize production workflows.

Problem 30: Carbon Emissions Tracking

Scenario: A large corporation needs to track Scope 1, 2, and 3 carbon emissions across their operations and supply chain.

AI / ML Specific (Problems 31-40)

Problem 31: Enterprise RAG System

Scenario: A law firm has 10M legal documents. Lawyers need to query them using natural language and get accurate, cited answers.

Key questions to ask:

• Document types? (PDFs, emails, contracts, case law)
• Accuracy requirements? (Legal context = very high)
• Latency? (Interactive search vs. batch analysis)

Approach:

• Sub-problems: (1) Document ingestion and chunking, (2) Embedding generation, (3) Vector store with metadata filtering, (4) Retrieval pipeline with re-ranking, (5) LLM generation with citations, (6) Evaluation and feedback
• Key trade-offs: Chunk size vs. context preservation. Speed vs. accuracy. Cost of LLM calls.
• Start with: 1,000 documents, one practice area, measure retrieval quality before scaling.

Problem 32: AI Agent for Customer Support

Scenario: A SaaS company handles 50K support tickets monthly. Build an AI agent that resolves 60% automatically.

Problem 33: LLM-Powered Data Analyst

Scenario: A business intelligence team wants non-technical users to query data using natural language.

Problem 34: Content Moderation at Scale

Scenario: A social platform needs to moderate 1M posts daily for harmful content.

Problem 35: Multi-Modal Search Engine

Scenario: A media company has 5M images, videos, and documents. Build a search system that accepts text, image, or audio queries.

Cross-Functional (Problems 36-50)

Problem 36: Data Migration Strategy

Scenario: Migrate a Fortune 500's data from Oracle + Hadoop to a modern cloud platform with zero downtime.

Problem 37: Real-Time Recommendation System

Scenario: A streaming service wants personalized recommendations updated in real-time as users browse.

Problem 38: IoT Data Platform

Scenario: A smart building company has 100K sensors across 500 buildings. Build a platform for real-time monitoring and analytics.

Problem 39: Compliance Monitoring System

Scenario: A financial institution needs automated monitoring of 500+ regulatory requirements.

Problem 40: Multi-Tenant SaaS Customization

Scenario: Your product serves 200 enterprise customers, each wanting custom workflows. Design a customization layer.

Problem 41: Event-Driven Architecture Migration

Scenario: Migrate a monolithic batch-processing system to event-driven real-time architecture.

Problem 42: Data Quality Monitoring

Scenario: A data platform has 10,000 tables. Build automated data quality checks with alerting.

Problem 43: API Gateway Design

Scenario: A company has 50 microservices. Design an API gateway for external partner access with rate limiting, auth, and versioning.

Problem 44: Search Infrastructure

Scenario: An e-commerce site with 10M products needs search that handles typos, synonyms, and personalized ranking.

Problem 45: Real-Time Dashboard for Operations

Scenario: An operations team needs a dashboard showing real-time metrics from 20 different data sources with <5 second latency.

Problem 46: Customer Data Platform

Scenario: Unify customer data from CRM, website, mobile app, support tickets, and purchase history into a single customer view.

Problem 47: Feature Store Design

Scenario: An ML team has 50 models in production. They're duplicating feature computation. Design a shared feature store.

Problem 48: Data Marketplace

Scenario: A data company wants to let customers discover, preview, and subscribe to datasets through a self-service marketplace.

Problem 49: Workflow Automation Platform

Scenario: A consulting firm has 200 consultants doing repetitive data processing tasks. Build a no-code/low-code automation platform.

Problem 50: AI-Powered Document Processing

Scenario: An insurance company processes 100K claims documents monthly. 80% are still manually reviewed. Automate extraction and classification.

Practice Tips

1. Time yourself. 45 minutes per problem. If you can't structure an approach in 5 minutes, your framework needs work.
2. Draw diagrams. Interviewers want to see visual thinking. Practice on a whiteboard or drawing tool.
3. Talk through trade-offs. There's no single right answer. Show that you understand the implications of your choices.
4. Ask questions first. The best FDE candidates spend 20% of the time clarifying the problem.
5. Start with the simplest version. Deploy a POC in week 1, iterate based on feedback.

Want to discuss specific solutions? Pick a problem number and post your approach in the replies. Community feedback is the best interview prep.