Scenario library
Golden scenarios
Every scenario has a written rubric, the failure modes an interviewer will inject, and a three-round format. Free includes 3 of them, one 5-minute mock each. The rest of the catalog, retakes, and untimed interviews are Pro. Sign in to start a timed mock.
32 scenarios
E-commerce Support Agent with Refund Actions (2M chats/month)
An agent that issues real refunds: idempotent tool calls, server-side policy limits, and handoffs humans don't redo.
Global E-commerce Rate Limiter (500k QPS)
500k QPS admission control with sub-2ms overhead, hotkey skew, and partition behavior.
Multi-Tenant AI Agent Gateway (50k RPM)
Per-tenant token buckets, provider fallback on 429/503, and an exact credit ledger at 50k RPM.
Billion-Scale Vector Retrieval for RAG
p99 under 50ms over 1B embeddings with ACL filters, one-minute freshness, and model migration.
Consumer Chat Assistant at 100M Weekly Users
Demand outruns a fixed GPU fleet: tiered admission, graceful model downgrade, and durable streams for 100M users.
Similar prompts reported at OpenAI
Cost-Controlled Social Intent Pipeline
Stream ingestion of 10M posts/day with tiered filtering and hard LLM budget governors.
Deep Research Agent with Parallel Subagents (50k reports/day)
20-minute agent runs that fan out to subagents: bounded loops and spend, crash-safe state, and citations you can trust.
Distributed Job Scheduler (100M jobs/day)
100M delayed and cron jobs per day with leases, fairness, retries, and midnight herds.
Double-Entry Payment Ledger (20k TPS)
Exactly-once double-entry transfers at 20k TPS with hot accounts and cent-level reconciliation.
Enterprise Knowledge Assistant with Source-System ACLs (8M docs)
Cited answers over wikis, tickets and Slack where a revoked permission must vanish from answers within minutes.
Similar prompts reported at OpenAI
Eval-Gated Model Rollout with Shadow Traffic (Weekly Releases, 30M DAU)
Ship a new model weekly without silent quality or safety regressions: eval gates, shadow diffing, staged canaries, fast rollback.
Similar prompts reported at Anthropic
Fault-Tolerant LLM Pretraining on 4,096 GPUs
At 4k GPUs something breaks every few hours: async checkpoints, hang detection and deterministic resume decide goodput.
Similar prompts reported at Anthropic
Flash Sale Inventory Pipeline (200k writes)
Zero-oversell inventory deduction under 200k concurrent writes with reservation rollback.
Invoice and Contract Extraction Pipeline (400k Docs/Day)
A wrong amount gets paid: schema-constrained extraction, calibrated per-field confidence, review routing and idempotent posting.
LLM Evaluation Platform Gating 20 Changes a Week
Ship-or-block verdicts in under an hour: calibrated LLM judges, per-slice statistics, and canaries that catch what evals miss.
Long-Term Memory for a Personal AI Assistant (20M users)
Remember the right facts across chats, resolve contradictions, honor deletes everywhere, and resist memory poisoning.
Multi-Model LLM Inference Serving Platform
Continuous batching, paged KV cache, and SLO-aware scheduling for 5k concurrent streams.
Multi-Tenant Business Chatbot Platform (20k Tenants, 15k Concurrent Chats)
Every business gets its own bot: tenant-isolated retrieval and caches, per-tenant quotas, and no noisy neighbor starving shared GPUs.
Multi-Tenant Fine-Tuning Platform (10k Concurrent LoRA Jobs)
Thousands of customer LoRA jobs share one GPU pool: fair scheduling, data isolation, safety eval gates and dense adapter serving.
Similar prompts reported at OpenAI
Multimodal Search over Images, Video and Text (540M Assets)
One query, three modalities: a shared embedding space, per-modality indexes, calibrated fusion and collapse of near-identical variants.
On-Device Phone Assistant with Cloud Escalation (500M Devices)
A 3B model in under 2 GB of RAM: quantization, local retrieval, privacy-preserving escalation and safe OTA updates at fleet scale.
Online Feature Store for Real-Time ML
p99 under 10ms feature serving with point-in-time correctness and training/serving skew control.
Real-Time Chat & Presence at 10M Connections
10M WebSockets, per-conversation ordering, large-group fanout, presence, and reconnect storms.
Real-Time Short-Video Recommendations (200M DAU, 150ms)
Multi-stage retrieval and ranking under 150ms, with swipes reflected on the next page and new uploads ranked within minutes.
Similar prompts reported at Meta
Real-Time Voice Support Agent with Barge-In (5k concurrent calls)
Sub-second voice-to-voice over noisy phone lines, interruptions that keep state honest, and no dead air when a stage stalls.
Repo-Aware AI Coding Assistant and Agent (8k engineers)
Sub-300ms inline completions, a fresh index of a 25M-line monorepo, and an agent that runs code without leaking secrets.
RLHF Post-Training Pipeline (70B Policy, 1M Preference Pairs/Quarter)
Reward models get gamed: rater quality ops, KL-bounded RL on a four-model GPU footprint, and human eval as the final arbiter.
Stateful Agentic Workflow Engine
Durable agent workflows with sandboxed tools, checkpoints, and human approval gates.
Streaming Guardrails Service for a Consumer Assistant (60k RPS)
Input and output filters inside a tight latency budget, stream-safe moderation, and policy changes that ship without a model redeploy.
Text-to-SQL Analytics Agent over a 12k-Table Warehouse
Find the right 10 of 12,000 tables, keep scans cheap, and never ship a plausible-but-wrong number to an exec.
Tiered AI Content Moderation (15M items/day)
The LLM can afford to see 5% of 15M daily items: cascade economics, per-category thresholds, appeals, and adversaries.
Zero-Hallucination Financial GraphRAG
Retrieval with citation grounding and deterministic arithmetic checks so numbers are never invented.