Two things in one place: real interview experiences and compensation packagesfrom ML loops at frontier labs — shared by people who’ve been through them. Sensitive numbers are members-only.
A four-round Research Scientist onsite at Scale AI that ended in a verbal offer. The first round was the toughest; by the end it leaned customer-facing. On the paper question people always ask: I do have a publication, but it wasn't weighted heavily.
Walked into a Harvey phone cold after years away from LeetCode — a build-up implementation problem with a formulas/references extension. Several companies apparently reuse this exact one, so it's worth having ready.
A three-round Harvey onsite. The coding round went fine, but the system design was LLM-based and I prepped for the wrong classic design — spent too long on RAG and missed the point. Reject a week later, almost certainly on the SD.
A Glean phone screen that was a hands-on ML-implementation problem in three parts. Passed — posting the general shape in case it helps someone prep.
A Datadog phone screen for the AI-agent role — a known coding problem plus a system-design piece (system design in a phone screen seems specific to the AI role). Passed; onsite in two weeks.
I barely passed Anthropic's phone coding round, then got a 40-minute hiring-manager round that was all classic behavioral. One note on a concurrency prep hint that never actually came up. Still waiting on the result.
A Snowflake backend new-grad onsite on the harder side — the phone problem paired a tree/forest structure with a little message-passing twist. Logging it; a few of the constraints felt underspecified in the moment.
An Okta Staff loop — a project deep-dive plus a LeetCode-style warmup on the phone, then a meatier Java low-level-design onsite that kept getting extended into concurrency and thread-safety. Passed.
A Staff-level Rippling loop that was pure system design, no coding, with three areas to cover and no room to go deep on any of them. Reject came next day — and from the other posts, it looks like most people don't clear this one.
Palo Alto Networks — the interviews themselves were on the easy side, but the whole thing fell apart at the end: no headcount approval, a promised team-match, then radio silence. More of a process story than a hard-questions one.
A phone screen for Sierra's Agent Engineer role — posting mostly to compare notes on what to expect next, since there's not much out there yet, and I'm still waiting to hear back.
A Meta phone screen — a string/ordering problem plus a two-pointer follow-up. Also genuinely unclear whether Meta had really un-frozen, so I'm posting to compare notes on the current req situation.
A Harvey onsite where the design round was an LLM/RAG-style build, and the real challenge was reading what they actually wanted me to optimize for. Good vibe overall; logging the shape of the loop.
Two phone rounds at Whatnot after the recruiter call — one coding, one system design — and none of it overlapped with what's already floating around here. Felt middle-of-the-road, and I ended up waitlisted.
A recruiter reached out and I had about two weeks to prep, which wasn't enough — the frequent questions really are that frequent and the bar is high. Rejected within two days. The thread turned into a useful prep-and-cooldown discussion.
A referral phone screen for an Amazon AI team — almost entirely an ML deep-dive, and they went deep on recent RL-for-LLM work rather than any coding. Passed. Heads up: every Amazon round carries a behavioral chunk too.
Two phone rounds at Clay (startup) through a recruiter — a system-design round that went great, then a coding round, then a quiet reject with no real feedback. My honest read: I was the backup and someone else took the offer.
A single Google L5 coding phone screen that leaned far more on Java async plumbing than on the algorithm itself — I walked out unsure and never got a clear verdict. Posting the setup, because a lot of people found the format surprisingly brutal for a phone round.
applied online for the FDSE (Forward Deployed Software Engineer) role. About two weeks later, HR reached out and scheduled the first-round phone interview, which was around 30 minutes. I’d recommend preparing well for the first round because they seem to take it pretty seriously. The interviewer asked what I knew about Palantir, why I was interested in Palantir, and why I wanted to join the company. Basically, make sure you can explain your motivation clearly and show genuine enthusiasm.
Just wrapped the OpenAI loop and it was a pass. Structure was a phone screen (one coding, one design) then an onsite with a chunky coding problem and a design round — no separate team-match step for me. I prepped almost entirely with hack2hire and it lined up well with what they asked.
All-whiteboard screen with a quant developer, zero IDE. It looked like a warm-up search problem but the whole thing was really about whether one API is expressive enough. A headhunter reached out to me for this, C++ background.
Onsite coding round and I didn't fully close out the follow-up, so I'm a bit in limbo. Trying to gauge whether my rubric is survivable: I think I'm sitting around 2 hire + 1 lean-hire + 1 hire/strong-hire.
Technical phone for the USDS org and honestly I just didn't click with it — I never even caught the interviewers' names, and I came out of it flat. Ended up a fail.
Phone-screen coding that leaned way more on explaining your design and walking examples than on racing to a solution. The thing that stood out: the interview was AI-assisted by default — they didn't even ask, the tooling was just there.
Senior SWE loop. The OA was a genuine grab-bag — multiple-choice across both SWE and AI, a DSA problem, an API-build problem, and even an optional prompt-engineering one. Round 2 was a system design plus a real deep-dive on past project decisions.
Two rounds stacked on the same day, back to back: a coding round then a system-design round. Came out with a pass. I'm now trying to figure out what Snowflake IC4 actually maps to at Google/Meta levels.
Technical phone and it ended in a fail. The main problem felt fresh — not something I'd drilled — and the comments actually argued about whether it even needs an algorithm or is just careful iteration. There was also a quick ML-fundamentals question.
Onsite, and this one stung: I got rejected even after the interviewer said, mid-round, "I'll buy you this design" — so I genuinely don't know what they were still missing. My one piece of advice is to actually write the thing out, because it's harder to implement cleanly than it looks.
Took this loop mostly as practice and still didn't pass, so grain of salt. The coding was a pretty ordinary OOD; the design round was the part that threw me because it went somewhere I hadn't prepped.
I booked what I thought was a routine recruiter chat and instead got dropped straight into ML questions with zero warm-up. The whole thing felt AI-mediated on both ends, and it made me realize the buzzwords on your resume can genuinely backfire now.
In-person onsite and, refreshingly, it was a good experience — food, coffee, decent energy, and no dreaded bar-raiser "dog" round for me. Sharing mostly because the in-person format made a real difference vs. virtual.
Passed the technical phone. Two medium LC-style problems, and the thing that clearly mattered was nailing every complexity follow-up. Side note that came up in comments: people are asking whether Uber is frozen right now (loops finishing with no team attached).
Full L2 loop and it was a gut-punch: solid marks on most rounds, but the AI-coding round I only got one question done, then I got ghosted on an HR call, had my references (two managers) called, and was told it was "insufficient signal" and rejected the next day as a final hiring-committee call — welcome to reapply in a year.
The BQ round with a veteran interviewer (IIT, 10+ years at Google), 3-4 "tell me about a time" questions, and it wrapped early. I'm nervous this is my weak link because the coding felt solid. Comments were reassuring — people say BQ rounds rarely sink candidates.
Senior-level loop and the surprise was there's no system-design round at all. Round 1 was a 30-minute hiring-manager conversation. I'm now on a waitlist and trying to figure out how long people usually wait after clearing the screen.
Phone screen that kicked off with C++ trivia before getting into the coding (which was paywalled on the source, so I can only share the opener).
Intern virtual onsite: two coding rounds back to back, coding only. Someone asked what "back to back" implied — it just means two straight hours with no special meaning, no SD round.
SDE3 loop, didn't make it through. The phone was an add-two-numbers variant, and the low-level design was a genuinely fun one about auto-assigning the nearest stocked machine to a user.
Sharing the process shape more than any single question, since that's what I wish I'd known. HR call first, then R1 blended algorithm and model-fundamentals questions with behavioral. The framing they liked: answer conclusion-first, then details, then a summary, and quantify your project impact.
A non-SWE-flavored loop that surprised people: heavy on culture fit plus a data-analysis task that needed some cloud-capacity intuition, and some SQL — but no traditional coding or system-design round. Commenters kept asking what role this even was.
Got Anthropic's newer interview format where the entire point is writing and reviewing code using AI tools — tool access is provided and you don't need to set anything up in advance. I applied through the online application; best guess is it's an infra role. Sharing the actual prompt they sent since a few people hadn't seen this format.
Referral phone screen. The core was a "rover" problem close to common practice questions, and then the interviewer pushed the scale hard — go from handling a few rovers to 1M+.
Rejected three business days later. What's frustrating is I did get a one-pass solution — but a simple bug ate too much of my time, so I never reached the follow-up, and I'm pretty sure that's why it was a no.
MLE phone screen, ended in a fail. I got a lower-frequency build-order problem, and since older posts about it were unclear, I'm writing the actual statement out so the next person isn't guessing.
Technical phone, ended in a fail. The specific tree problem was behind the paywall on the source, so I can only log it for the record rather than reproduce it.
Referral technical phone. It combined a full URL-shortener design (both high- and low-level) with a find-the-bugs exercise. Details were partly paywalled but the shape is clear.
PhD-track loop. I designed a KV cache in Java, spent time debating whether the value could be an int or a string, and then ran out of clock on the transactions part. Half-jokingly: Python would've been comfier for this.
3.5 YOE, C++ role. The phone was implementing std::vector from scratch — I finished in about 45 of the 90 minutes and we chatted the rest. Broader context: I've been searching ~4 months, done 5-6 final rounds, and have 0 offers yet, so I'm a bit worn down.
MLE/Research PhD loop: two research-design rounds (open scenarios plus language-model topics, including a linear-attention discussion) and one behavioral round. The recruiter seemed specifically dedicated to the MSL/MRS tracks.
Passed! About 50 minutes, mostly problem-solving, all test cases passing, and a next-steps email within 30 minutes of hanging up. It looks like a junior-leaning req — the VO structure was laptop + SD + HM.
Intern OA, 75 minutes, and it was three quite different things rather than one algorithm grind. New grad here just trying to land something.
Intern OA, 120 minutes, a single question but a meaty one. Moderate difficulty. New grad grinding for an offer.
Intern VO after the OA. STAR behavioral, real Python questions, a walkthrough of my OA code, and a money/precision discussion — with follow-ups that make you modify code on the spot.
Full-time via headhunter. Friendly Chinese interviewer, detailed resume questions, then LC295 and a map-vs-unordered_map discussion. Didn't pass, but it was a pleasant round.
ML-Eng PhD loop, DL coding round. They asked about Focal Loss and then to implement it by hand. Hard, and I didn't pass.
The hardest screen I've done. You're handed a real security-event-processing codebase and asked to review it, apply a design pattern, and implement. Failed it, but it was a legitimately good round.
SDE2 loop. The OA had two DSA questions (paywalled), and the video interview was AWS services plus a walk through my resume experience. Didn't pass.
New-grad QR-ML screen: a 40-minute phone with a quant researcher and one modeling problem (paywalled). It leaned on probability and multi-step reasoning more than coding.
Easy-rated 60-minute phone and it was a pass. One main problem plus a follow-up, the design was friendly, and the interviewer gave off a positive read the whole time. Someone asked SDE vs MLE — this was the general track.
Passed! A 75-minute Staff+ design round where they explicitly let you pick a traditional approach or an LLM-based one. The curveball was a latency-estimation question I had to answer fast.
Passed, after a scare. 75-minute Staff+ design: they let you choose a traditional or LLM approach (I went LLM), and there was also a sales-data-stream problem. The interviewer screenshotted my design as we went, which made me think I'd bombed — then the yes came 4 days later.
AI-voice startup (Chinese founder, several Chinese engineers on the team). Round 1 was a system design: a system that processes various SQL syntax, and the interviewer kept adding features so I was constantly re-scoping. No leetcode.
CS PhD, 5 YOE, coming from a Data Scientist/MLE role at a traditional midwest company. Two OA-style rounds and then a one-year HR freeze after the reject, which I'm still salty about. R1 was science depth; R2 a coding problem I ran out of time on.
Passed the MTS loop — it was rated the hardest tier but it was doable. The DSA round had a topological-sort classic plus a prime-splitting counting problem, and the system-design round was a parking/car-rental booking system.
Meta's newer AI-assisted coding format caught me off guard. It looks like one problem, but it's got staged, unannounced follow-ups you have to finish in the time limit — very different from the traditional "we'll tell you how many questions and give you a break" flow. Didn't pass.