Traverse rapid-fire + the RabbitMQ claim, handled

The drill set, then the script.

1. What is Traverse? → passive DSA capture (extension) + spaced-repetition scheduling + streaks/analytics, iOS app + backend.
2. How does capture work? → MAIN-world network interception in the extension; platform adapters normalize judge responses; POST with idempotency key + device id.
3. Why a queue? → 15 s LLM ladder off the request path; 10 s → 0.2 s p95 (Day 4).
4. Which queue, precisely? → BullMQ on Redis. (And the resume script — see below.)
5. How do you dedup submissions? → @@unique([userId, idempotencyKey]); duplicate POST returns 200 with the existing row.
6. What happens on the AI provider failing? → ladder falls through (BYOK Gemini → server Gemini → Groq → G4F), JSON cleaned/validated, heuristic fallback, scheduling never blocks on AI.
7. How does scheduling work? → FSRS-style memory state per problem/user-topic; due dates from retrievability targets (0.85); predictions persisted with model version; calibration measured.
8. What's the calibration harness? → predicted vs observed, bucketed by modelVersion (Day 72).
9. Free vs premium? → AI signals nulled at one choke point for free users; the deterministic scheduler path is always exercised.
10. How do notifications work? → intent (Notification) + per-channel ledger (NotificationDelivery) with lease claim, backoff, terminal statuses; push best-effort via APNs/FCM; in-app truth.
11. Why no WebSockets? → push for closed apps, polling for freshness, no stateful socket fleet (Day 73).
12. Cache strategy? → cache-aside, per-data TTLs (120–600 s), worker-driven invalidation, 2 s timeouts, degrade-to-no-op.
13. Redis policies? → noeviction (queue safety over cache elasticity), AOF, removeOnComplete bounds (Day 46).
14. What breaks at 2 replicas? → in-memory rate limiters and node-cron jobs (Day 50).
15. Biggest query perf win? → streak walk 300→2 queries; achievements 120→4 (Day 28).
16. Durability? → WAL archiving to S3, 5-min RPO, Aiven offsite, executable restore endpoints (Day 34).
17. Deploy? → GitLab CI → SSH → compose, git-reset + rebuild + health gate, migrate-on-boot refusing stale schemas.
18. Security posture? → bcrypt-12, strict auth limits, webhook HMAC over raw body, log redaction, BYOK isolation; debts: hardcoded provider key, two timing-unsafe compares (Day 90).
19. How do you know it works? → integration tests replaying real migrations, per-feature regression tests (the wiring bugs I found and guarded), calibration for the scheduler.
20. Social graph? → double-row friendship, canonical pair streaks with CHECK constraint, blocks filtered by a single visibility predicate.
21. What's the recon feature? → streaming unbounded capture dumps to disk, .partial + atomic rename, per-platform quota, digest email (Day 82).
22. Biggest scaling limit? → single VPS, single Redis, in-memory limiters; the ladder is Redis-backed limits → read replicas → partition submissions.
23. What would you rebuild? → submission outbox (close the enqueue gap), Redis split, token expiry on JWTs, encrypted BYOK keys.
24. Why Node/TypeScript? → one language across extension, backend, tooling; the workload is I/O-bound, not CPU-bound.
25. What are you proudest of? → the fact/derivation split that made a 50× latency win safe, and the calibration harness that measures the scheduler instead of asserting it works.

The RabbitMQ script (memorize verbatim)

If asked "your resume says RabbitMQ for Traverse":

"My resume summarizes the pattern as a message queue — the implementation is BullMQ over Redis, and RabbitMQ is the technology in my other project, Zoth. The architectural decision is identical in both: decouple ingestion from downstream processing so user-facing writes don't pay for slow side effects. In Traverse I chose a Redis-backed queue because the jobs are small triggers whose truth lives in Postgres, and a single VPS didn't need a second broker; in Zoth I chose RabbitMQ because provisioning events needed broker-native retry accounting, DLQ semantics, and process isolation — and I can go deep on either."

Then pivot to what you did build (the 50× win), not to the discrepancy. Own it, don't litigate it.