T1 · JUNIOR
Design semantic search for a docs site
Embeddings, a vector index, and a golden set. The retrieval warm-up.
30 min
PRACTICE BANK
Where AI system design gets real.
Tier 1 is a semantic search warm-up. Tier 4 is a multi-tenant GPU fleet on fire. Your tutor watches every whiteboard.
DONE0/ 25
IN PROGRESS—
UNLOCKED12for Mid
AVG. SESSION—min
BEST STREAK—days
Sort: Curated
T1FoundationsScoped single-pipeline problems. Prove you can name the parts.5 problems · 0 done
T1 · JUNIOR
Design smart compose (email autocomplete)
Sub-100ms LLM completions at keystroke rate. Latency is the product.
30 min
T1 · JUNIOR
Design a FAQ support chatbot
Grounded answers, guardrails, and a human handoff.
30 min
T1 · JUNIOR
Design a spam & toxicity classifier
Classic ML under class imbalance. Choose the right metric first.
25 min
T1 · JUNIOR
Design a meeting summarizer
Long transcripts, structured notes, factuality checks. Async by nature.
25 min
T2CommonThe interview canon. RAG, recsys and ranking at real scale.7 problems · 0 done
T2 · MID
Design 'chat with your docs' for an enterprise
The most-asked GenAI question. Retrieval quality, permissions, evals.
45 min
T2 · MID
Design video recommendations (TikTok 'For You')
The classic ML flagship. Two-stage pipeline, features, cold start.
45 min
T2 · MID
Design fraud detection (Stripe)
An adversary in the loop. Fresh features, skew, delayed labels.
40 min
T2 · MID
Design ad click prediction
The Meta staple. Features, leakage, calibrated probabilities, A/B.
40 min
T2 · MID
Design AI-powered product search
Hybrid retrieval, rerank, generated answers. Search that cites.
40 min
T2 · MID
Design an ETA / delivery-time prediction system
The regression staple. Geospatial features, real-time traffic, freshness.
40 min
T2 · MID
Design a document extraction pipeline
PDFs to typed JSON at scale. Schemas, retries, error analysis.
35 min
T3AdvancedGPU budgets, eval platforms, agents — the senior deep dives.7 problems · 0 done
T3 · SENIORabove your band
Design ChatGPT
The flagship. Streaming, memory, moderation, GPU serving, cost.
60 min
T3 · SENIORabove your band
Design an AI coding assistant (Copilot)
Tab-completion latency + chat depth. Context is the scarce resource.
50 min
T3 · SENIORabove your band
Design an LLM evaluation platform
Judges, golden sets, CI gates. The system that grades the systems.
50 min
T3 · SENIORabove your band
Design an inference system for a single GPU fleet
Anthropic's actual question. Batching, KV memory, goodput under SLOs.
50 min
T3 · SENIORabove your band
Design visual search (Pinterest Lens)
Image embeddings at billions scale. Index choice, rerank, freshness.
45 min
T3 · SENIORabove your band
Design real-time recommendations
Session-aware recs in seconds. Streaming features, freshness SLAs.
50 min
T3 · SENIORabove your band
Design an agentic deep-research feature
Multi-step search agents that compress the web into an answer.
50 min
T4ExpertAmbiguous, org-scale, adversarial. Cost cuts and critiques.6 problems · 0 done
T4 · STAFFabove your band
Design a multi-tenant LLM inference platform
Many models, many teams, one GPU fleet. Routing, isolation, unit cost.
60 min
T4 · STAFFabove your band
Design an autonomous workflow agent
An agent that acts on real systems. Permissions, sandboxes, injection.
60 min
T4 · STAFFabove your band
Cut this AI system's cost 10×
A working product bleeding money. Find the multiplicative levers.
40 min
T4 · STAFFabove your band
Design an org-wide AI gateway
Every team, every model, one front door. Budgets, evals, governance.
55 min
T4 · STAFFabove your band
Design a text-to-image generation service
20-second renders, GPU queues, content safety, cost per image.
55 min
T4 · STAFFabove your band
Critique: find the failure modes in an AI system design
A plausible AI architecture is on the board. Find what breaks first.
45 min