Google · Phone screen → Onsite
In processGoogle SWE III ML Experienced Hires: 34 Experiences (Part 1 of 4)
Google SWE III (L4) ML-track loops for experienced hires, January to fall 2026: coding, an ML design or domain round and Googleyness, with offers at L4.
At a glance
- Process: a phone stage or GHA → two or three coding rounds → an ML domain or ML system design round → Googleyness & Leadership (G&L) → hiring committee and team matching (entries 01–10).
- Asked recently: ML design on harmful-video detection with embeddings (01), query-based document search (09) and an AI coding system (05); coding such as Russian Doll Envelopes (01) and lakes inside an island (09).
- Difficulty: coding ran medium to lower-hard (10); levels moved both ways, with L5 loops ending at L4 (03, 09) and L4 loops ending at L3 (06, 07, 08).
- Practise first: Rotate a flat image clockwise, Count lakes inside an island and video safety embeddings.
Coverage: Experienced hires only (stated work years, a current or past employer, or earlier full-time roles). Parts 1–3: Software Engineer III / L4 ML-track loops (titles like SWE III AI/ML, SWE-ML and MLE), a few starting at L5. Part 4: Forward Deployed Engineer and Cloud AI Engineer. US, Europe and India. This part: 10 loops from January to fall 2026.
In this series
- Part 1 (current): L4 ML-track loops from fall 2026 back to January 2026 (entries 01–10), plus the topic summary for all 28 ML-track reports.
- Part 2: ML-track loops from November 2025 back to June 2025 (entries 11–18), one with a Gemini team-match round.
- Part 3: ML-track loops from May 2025 back to March 2024 (entries 19–28), then patterns across the ML track.
- Part 4: Forward Deployed Engineer and Cloud AI Engineer loops.
Candidates who did not state their work experience have their own series: Google SWE III ML, hiring level not reported.
What came up
A behavioral round appears in 20 of the 28 ML-track reports, and 13 describe an ML design or applied-ML case. Graph problems and string problems were the most repeated coding themes, with five reports each.
| Topic | Reports | Last seen |
|---|---|---|
| Behavioral / Googleyness & Leadership round | 20 of 28 | Fall 2026 |
| ML system design or applied ML case | 13 of 28 | Fall 2026 |
| Arrays and sorting | 3 of 28 | Fall 2026 |
| Graphs, grids and shortest paths | 5 of 28 | Oct 2026 |
| Dynamic programming | 2 of 28 | Oct 2026 |
| ML theory and fundamentals Q&A | 5 of 28 | Jul 2026 (report date) |
| LLMs, attention and GenAI | 4 of 28 | Jun 2026 |
| Backtracking and constraint search | 2 of 28 | Jun 2026 |
| GPUs and hardware | 2 of 28 | Mar 2026 |
| Strings and words | 5 of 28 | Jan 2026 |
| Retrieval, ranking and recommendation | 3 of 28 | Jan 2026 |
| Coding with an ML twist | 2 of 28 | Jan 2026 |
| Hash maps, heaps and counting | 4 of 28 | Aug 2025 |
| Streams and sliding windows | 3 of 28 | Aug 2025 |
| Trees | 2 of 28 | May 2025 |
Binary search (#26) and statistics (#23) came up once each. A candidate can sit in several rows, so the counts don't add up to 28.
Questions we found
These tables cover the 10 reports in this part; Parts 2 and 3 have their own. Each question here comes from one candidate.
Algorithms / Data Structures
| Question | Reports | Last seen |
|---|---|---|
| Rotate a flat image clockwise, given as a 1D pixel array | 1 | Fall 2026 |
| Russian Doll Envelopes (LC 354), with rotatable envelopes | 1 | Fall 2026 |
| Grid DP problem in the Max Area of Island family, twist withheld | 1 | Sep 2026 |
| Constraint-satisfaction problem in the style of N-Queens | 1 | Jun 2026 |
| Minimum-cost path finding | 1 | Jun 2026 |
| String grouping, harder than Group Anagrams (practice version: group anagrams in first-seen order) | 1 | Jan 2026 |
| Count lakes inside an island | 1 | Jan 2026 |
ML Design & Modeling
| Question | Reports | Last seen |
|---|---|---|
| Detect harmful YouTube videos, centred on embeddings (learn video representations for safety review) | 1 | Fall 2026 |
| Design an AI coding system (related match: design a code-generation language model; exact prompt withheld) | 1 | Jun 2026 |
| Model for query-based document search | 1 | Jan 2026 |
| Binary classifier that adds navigable links to Google Search results (decide when a search result gets shortcut links) | 1 | Jan 2026 |
ML Theory & Fundamentals
| Question | Reports | Last seen |
|---|---|---|
| ML project lifecycle: metrics, features, model families, data prep | 1 | Jul 2026 |
Behavioral (BQ)
| Question | Reports | Last seen |
|---|---|---|
| Improving a process | 1 | Fall 2026 |
| Taking on too much at once | 1 | Fall 2026 |
| A stretch goal you missed | 1 | Fall 2026 |
| Working with shifting or unclear information | 1 | Fall 2026 |
| Ramping up outside your own specialty | 1 | Fall 2026 |
| Bringing in user and teammate perspectives | 1 | Fall 2026 |
10 detailed candidate timelines
G&L is Googleyness & Leadership, GHA is the Google hiring assessment, and HC is the hiring committee.
01: SWE AI/ML L4 loop with two coding rounds and a video-safety design
When: fall 2026 (report 2026-11-04; round order not given) · Level: L4 · Outcome: not reported
- Coding (45 min): rotate a flat image 90° clockwise; the input is a 1D pixel array plus the row and column counts.
- Coding (45 min): Russian Doll Envelopes (LC 354); follow-up: envelopes may be turned 90°. FastPrep's statement keeps envelopes fixed; only its source note mentions the turn.
- ML system design: detect harmful YouTube videos, centred on embeddings for harmful content; practice: learn video representations for safety review.
- G&L: six themes: a missed stretch goal, taking on too much, improving a process, ramping up outside one's specialty, shifting or unclear information, and user and teammate perspectives.
02: SWE-ML L4 loop that passed, then a startup offer
When: stage 1 ~2026-09-10; onsite the week of 2026-10-13 (three slots booked, two described) · Level: L4 · Where: Europe · Outcome: passed; took a competing startup offer while staying in the EU team-matching pool (2026-11-05)
- Stage 1, coding: a grid/graph DP problem in the Max Area of Island family, with a twist the candidate kept private.
- Stage 1, behavioral: standard questions.
- Onsite, coding 1: problem not named.
- Onsite, coding 2: problem not named.
03: L5 loop retargeted to L4 that stopped before the onsite
When: report 2026-09-16 (year inferred) · Level: L5, retargeted to L4 after the coding round · Outcome: not moved on to the in-person onsite
- Coding and Googleyness: no task details.
- AI domain (virtual; added because of a cloud/AI focus): not described.
04: SWE-ML L4 phone stage with an ML lifecycle round
When: report 2026-07-24 (year inferred; the phone stage came after posting, order not given) · Level: L4 · Outcome: later stages not reported
- Phone stage, Googleyness: not described.
- Phone stage, ML domain: an overview of an ML project's lifecycle: picking metrics, featurization, model families for classification and regression, and data preparation.
05: SWE-ML L4 loop with an AI-coding design and an in-person coding stage
When: GHA, then R1 before 2026-06-17 (format not stated); R2 in person 2026-06-18 · Level: L4 · Outcome: pending at last report
- R1, Googleyness: standard questions.
- R1, ML system design (picked over a DSA option): design an AI coding system, exact prompt withheld; related match: design a code-generation language model.
- R2, coding (in person): a constraint-satisfaction problem similar to N-Queens; inputs and outputs not reported.
- R2, coding (in person): minimum-cost path finding, with follow-ups.
06: L4 MLE loop that moved to L3 team matching
When: final round ~2026-05-18 (inferred from a 2026-05-28 results call; in person; round count and order not given) · Level: interviewed for L4 · Where: US · Outcome: accepted L3 team matching; a one-line reply on 2026-10-21 says L4 was settled after HC, with no later confirmation
- Coding: problems not named.
- ML round: a domain-specific ML design question, prompt not shared.
- Behavioral: standard questions.
07: SWE-ML loop that restarted with add-on rounds and ended at L3
When: first pass Sep–Nov 2025; add-on rounds Mar 2026 · Level: an L4/L3 role · Outcome: L3 offer through hiring review, May 2026
- GHA (Sep 2025).
- First round, phone interview (Oct 2025): not described.
- Final interviews (Nov 2025): composition not given.
- Add-on, GPU programming (Mar 2026): content not given.
- Add-on, technical (Mar 2026): not described.
The first pass was rejected at hiring review with a six-month cooldown; the hiring manager asked to restart in Feb 2026.
08: SWE III AI/ML loop down-leveled to L3
When: round 1 ~mid-Feb 2026; round 2 date not given · Level: interviewed for L4, down-leveled to L3 · Outcome: in L3 team matching since Mar 2026, still pending on 2026-07-14
- Round 1, ML domain (virtual): content not reported.
- Round 1, Googleyness (virtual): not described.
- Round 2, coding ×2 (onsite): problems not named.
HC asked for one more coding round; whether it was held isn't stated.
09: L5 SWE AI/ML loop that ended in an L4 offer
When: screening Dec 2025; online onsite ~2026-01-14 (round order not given); coding retake mid-January 2026 (a later recap says Feb) · Level: L5 loop; HC approved L4 · Where: Europe · Outcome: offer at L4, 2026-04-22
- Screening, coding (Dec 2025): not described.
- Googleyness: behavioral discussion.
- ML domain (search/LLM angle): a model for query-based document search.
- Coding (Google's code editor): count the lakes inside a given island, from the island's coordinates.
- ML system design (Google Docs): a binary classifier over queries that adds navigable links to Google Search results; practice: decide when a search result gets shortcut links.
- Coding retake: a string-grouping task harder than Group Anagrams, exact task not given; practice version: group anagrams in first-seen order.
The coding round was retaken after the editor failed; a later comment from the candidate also links the retake to the first round's score.
10: L4 SWE AI/ML loop with an ML Q&A round and an L4 offer
When: Nov 2025 to 2026-01-06 · Level: L4 · Where: US · Outcome: offer at L4, 2026-01-23
- Mock interview (2025-11-21): practice with a Googler.
- Round 1, ML domain (2025-12-08): Q&A on a topic the candidate chose (GenAI), with more Q&A than ML coding; questions under NDA.
- Round 1, G&L (2025-12-18): not described.
- Round 2, coding ×2 (2026-01-06, in person): medium to lower-hard problems from Google's own question bank.
Two team-match calls on 2026-01-12, match confirmed 01-13, HC passed 01-21.
Common questions
What is the Google SWE III ML domain round?
Either an ML system design (harmful-video detection in 01, an AI coding system in 05, search shortcut links in 09) or a Q&A on ML topics (a project lifecycle in 04, GenAI in 10).
How many coding rounds are in the Google L4 ML loop?
Usually two, sometimes after a coding phone stage (01, 02, 05, 08, 10); one candidate retook a coding round after the editor failed (09).
Can Google change your level after the loop?
Yes: two L5 loops ended at L4 (03, 09) and three L4 candidates moved to L3 matching or an L3 offer (06, 07, 08).
How long does Google team matching take after HC?
From days to months: entry 10 matched a day after two calls and had HC pass about a week later, while entry 08 was still in L3 matching four months on.
Keep practicing
- Coding from this part: rotate a flat image and Russian Doll Envelopes (01); lakes inside an island and anagram groups in first-seen order (09).
- ML cases from this part: video safety embeddings (01), code-generation model design (05), query-document search and search shortcut links (09).
- Google question banks: algorithms and system design (sign-in needed).
- Everything Google in one place: the Google interview page and Google hiring insights.
- More firsthand reports: the Experience feed.
In this series
Sources: Anywhere useful interview info can be found.
Comments
0No comments yet — be the first to share what you know.