System design problems answered as 60-minute interviews, then stress-tested in an adversarial interview round (Codex as interviewer) and grounded against published industry practice with verified sources.
The 90-minute build session: direct an AI to a tested Go rate limiter — CSV replay, oracle-verified core, web server — then defend it for production.
Git push to running service: build isolation, image caching, deploy state machine, zero-downtime rollout, rollback, and rapid successive pushes.
Durable background jobs and workflows: leases, retries, checkpointing, fan-out, cancellation, and per-tenant fairness at 500M executions/day.
10-50 GB/s of multi-tenant logs: backpressure, live tail in 2s, tenant isolation, honest dropped-log accounting, hot/cold tiers.
Platform subdomains and 1M custom domains to the right container: edge route distribution, TLS at scale, surviving control-plane outages.
One-click Postgres with sub-60s failover: replication modes, split-brain fencing, stable connection URLs, PITR, and the whale-customer problem.
3M containers on 50k hosts: bin packing, reconciliation, noisy neighbors, persistent-disk pinning, scale-to-zero cold starts, region evacuation.
Versioned env config and secrets with atomic snapshots, shared-group fan-out blast radius, plus reliable deploy webhooks with replay.
S3-compatible storage at 10 PB: erasure coding, metadata and listing, multipart, 11-nines durability math, run by a small team.
Identity for humans, workloads, and AI agents: SSO/SCIM, short-lived scoped credentials, sub-10ms authz, and migrating away from leaky API keys.
Traces for agentic apps: OTel GenAI spans, prompt/completion payload handling, billable token attribution, 2-second live debugging.
An MCP server that lets agents operate infrastructure: risk tiers, plan-then-apply, prompt-injection defenses, bounded blast radius.
Metering to invoice: late/duplicate events, asymmetric accuracy, spend caps enforced in minutes, and an audit trail from raw event to line item.
Enterprise search and RAG where every result must respect source-system ACLs: filter coarsely in the index, check authoritatively at read time.
1B events/day delivered to flaky customer endpoints: dual-region ingest, per-endpoint isolation, tiered retry topics, ClickHouse attempt history.
Multi-tenant CI at ~12k job starts/s: Firecracker cells, lease-based failure detection, fair-share scheduling, petabyte-scale artifacts.
100M payments/day with a double-entry ledger at the center; ambiguous processor outcomes become an explicit UNKNOWN state that reconciliation resolves.
A web playground for streaming LLM inference: SSE streams that survive disconnects, spend caps, idempotent runs, org-sharded storage.
Team chat shipped in two weeks on Phoenix/Postgres, then the road to 10B messages/day with channel servers and ScyllaDB.
Video generation on a scarce GPU fleet: admission control in GPU-seconds, gang scheduling, preemption tiers, checkpointed long jobs.
Traces in, verdicts out: an eval platform with LLM-as-judge controls, holdouts against overfitting, and regression gates on releases.
Global live streaming with on-demand replay: a media plane that survives control-plane loss, epoch-fenced ingest, token-authed segments.
An LLM support agent with tool access: retrieval, action execution, escalation to humans, and the guardrails around all three.