Changelog

Robinhood joins the catalog: 46 questions reconstructed from 94 scraped threads spanning January 2025 to October 2026 — the densest company on the source forum so far. The anchor is the distributed job scheduler, reported more than 3x anything else and sometimes named on the phone-screen invite itself; all seven system-design questions ship with full breakdowns, and every coding question reported three or more times has a solution executed against its own test cases. Design rounds here grade correctness under failure— at-most-once execution, cancel-versus-fill races, worker crashes — and coding rounds run long: reports estimate 150–200 lines, graded on working code over elegant structure. The free preview is the service load-factor DAG round (frequency 14) with its follow-up ladder and tested solution. Included with any subscription.

Zip (ZipHQ) joins the catalog: 17 questions reconstructed from 33 candidate reports spanning February 2025 to July 2026, twelve of them with solutions executed against their own test cases. The loop is front-loaded: reports show candidates rejected at the phone screen more often than at the onsite, and there are frequently two phone screens — a LeetCode-style round, then a practical round inside an existing repository where you fill in business logic and write its unit tests. That practical round is the most-reported question here. The self-serve path is a 90-minute, four-level gated OA where each level’s tests must fully pass before the next unlocks; one candidate cleared three of four, scored 750/1000, and was still rejected. The free preview is an LRU cache debugging challenge whose two planted bugs cancel each other on the obvious test — fix only one and a passing case starts failing. Included with any subscription.

Roblox joins the catalog: 34 questionsreconstructed from 200+ candidate reports across 135 threads spanning February 2025 to August 2026 — the largest corpus here after the two AI labs, and the first non-AI company. Roblox’s question set is unusually concentrated: a sliding-window rate limiter at 29 reports and a most-frequent call-path problem at 22 account for roughly a third of everything reported. Both arrive as graded parts submitted one at a time, and in both the base problem is not what decides the round — the follow-up is. Two separate reports describe passing every visible test on the rate limiter and failing a hidden one for the same reason. The design round is most often like/unlike at scale or delayed payments, and reports reward naming the Redis key and the read and write queries over staying at box-and-arrow level. Included with any subscription.

Harvey AI joins the catalog: 9 questions reconstructed from 50+ candidate reports spanning March 2025 to August 2026. Harvey repeats itself more than any other company covered here — twelve of the fifty reports describe the question they were given as the known one from the forum. The phone screen draws from a small rotating set, most often the RAG citation / text-matching round, which at 23 reports is among the most-reported questions in the entire catalog; the onsite design round is almost always Google Drive-style storage with ACL or a RAG pipeline. Included with any subscription.

Palantir joins the catalog: 18 questions reconstructed from 28+ candidate reports spanning January 2025 to July 2026 — including the two rounds nothing else in the catalog covers: the Learning round, where the interviewer teaches you an unfamiliar system (a single-threaded package manager for SWE, Palantir’s internal SQL-like query language for FDSE) and asks you to extend it, and the open-ended Decomposition round. Candidate reports state directly that Palantir reuses its questions and splits them by track. Included with any subscription.

Mercor joins the catalog: 3 questions reconstructed from 14+ candidate reports — the distributed job-scheduler system-design round (seen for both Marketplace and Frontier Data MLE roles), the signature 3-part referral-network coding round with a harness-tested solution, and the faulty-comparator top-k algorithm round. Every Mercor question is free — no account required. More Mercor questions ship as new reports come in.

Perplexity joins the catalog: 8 questionsreconstructed from 29+ candidate reports — the OA-style multi-part coding rounds Perplexity is known for (in-memory Unix file system, an “Implement ToDo list for AI” task DAG with cascade failure, the stop-words streaming problem), the byte-tokenizer round, beam search, an expiring-credit tracker and a temporal key-value store with restore, plus a codex-fact-checked Design Mint.com system-design walkthrough, all with harness-tested solutions. The in-memory file system question is free — no account required.

xAI joins the catalog: 16 questions reconstructed from 80+ candidate reports — the staged in-memory database with nested transactions (reported 9 times), the complete 15-minute phone-screen script (reported 10 times), a lazy-refill rate-limiter walkthrough, the RAG pipeline design-plus-evaluation round, GPU node failure detection, a concurrent job scheduler, and harness-tested coding solutions throughout. The phone-screen guide and the in-memory database question are free — no account required.

Waymo joins the catalog: 28 questions reconstructed from 55+ candidate reports — real technical phone screens (BFS/graph, pure-Python data processing), AV-flavored system design (evaluation systems, inference at 100M DAU), six coding solutions, and full system-design walkthroughs. The entire Waymo catalog is free — no account required.