Google · Phone screen → Onsite
In processGoogle SWE III ML Interviews: 11 Experiences (Part 1 of 3)
Google SWE III / L4 ML-track interviews, hiring level not reported: 11 loops from Oct 2025 to Nov 2026, mixing coding, ML design and behavioral rounds.
At a glance
- Process: a phone screen or online coding, then a virtual loop of two or three coding rounds, an ML or AI domain round and Googleyness & Leadership (04, 07, 08, 09, 10). The hiring committee added extra rounds in 04, 06 and 09, and team matching can stall (07, 08).
- Asked recently: graph problems in 3 of the 11 reports here (04, 05, 10), DP in 2 (04, 05), and ML design or applied-ML cases in 4 (02, 05, 10, 11).
- Difficulty: report 04 rated the onsite coding LeetCode medium and the add-on rounds medium-to-hard and hard; every round in 10 ran 45 minutes.
- Practise first: Minimize the Largest Edge on a Path, Top-Rated Movie Among Transitively Similar Movies and a Google Reviews-like ML design.
Coverage: hiring level not reported (no years of experience or employer given). Parts 1–2: Software Engineer III / L4 ML-track loops (SWE III ML, SWE AI/ML, MLE). Part 3: Forward Deployed Engineer and Cloud AI Engineer, which are different jobs. Locations named include the US, EU/EMEA and India; interviews from 2024 to fall 2026. Applied AI appears once, as a team at team matching (01).
In this series
- Part 1 (current): ML-track loops from November 2026 back to mid-October 2025 (01–11), plus the topic summary for all 21 ML-track reports.
- Part 2: ML-track loops from early October 2025 back to April 2024 (12–21), then patterns across the ML track.
- Part 3: Forward Deployed Engineer and Cloud AI Engineer loops, each in its own role section.
Candidates who did state their work experience are in a separate series: Google SWE III ML, FDE and Cloud AI interviews, experienced hires.
What came up
Across all 21 ML-track reports in Parts 1–2, a behavioral round shows up in 16 and an ML design or applied-ML case in 8. Graphs were the most repeated coding theme. One candidate can sit in several rows.
| Topic | Reports | Last seen |
|---|---|---|
| Googleyness & Leadership (behavioral) | 16 of 21 | Sep 2026 |
| ML system design or applied-ML case | 8 of 21 | Sep 2026 |
| Graphs, grids and shortest paths | 5 of 21 | Jun 2026 |
| ML theory and fundamentals | 4 of 21 | Jun 2026 |
| Dynamic programming | 3 of 21 | Jun 2026 |
| LLMs, attention and GenAI | 3 of 21 | Apr 2026 (report date) |
| Strings and words | 2 of 21 | Sep 2026 |
| Hash maps, heaps and counting | 2 of 21 | Apr 2026 (report date) |
| Coding with an ML twist | 2 of 21 | Sep 2025 |
Single reports covered trees and a non-ML system design (02), binary search (10), statistics (18) and backtracking (21).
Questions we found
These tables cover the 11 reports in this part; Part 2 has its own. "Title only" means the candidate named the question but gave no details.
Algorithms / Data Structures
| Question | Reports | Last seen |
|---|---|---|
| Decode String with tree traversal (title only; practice: decode a bit string with a code tree) | 1 | Sep 2026 |
| Top-rated movie among transitively similar movies | 1 | Apr 2026 (report date) |
| Count subsets whose sum is above half the total | 1 | Apr 2026 (report date) |
| Minimize the largest edge on a path | 1 | ~Dec 2025 |
| Shortest path between two nodes | 1 | ~Dec 2025 |
| Binary search on the answer (task not named) | 1 | ~Dec 2025 |
System Design
| Question | Reports | Last seen |
|---|---|---|
| Design a distributed logging system (title only; filed under centralized log ingestion and search) | 1 | Sep 2026 |
ML Design & Modeling
| Question | Reports | Last seen |
|---|---|---|
| A Google Reviews-like system | 1 | Apr 2026 (report date) |
| Metrics for imbalanced data (related practice) | 1 | ~Dec 2025 |
| Deploying and monitoring a model (related practice) | 1 | ~Dec 2025 |
| Debugging a weak model | 1 | ~Dec 2025 |
| An image-classification app | 1 | Oct 2025 |
ML Theory & Fundamentals
| Question | Reports | Last seen |
|---|---|---|
| Résumé-driven ML theory: hypotheticals, implementation, metrics | 1 | Jun 2026 |
| ML fundamentals why-questions (bias-variance, regularization, clustering, entropy, time series) | 1 | ~Dec 2025 |
Behavioral (BQ)
| Question | Reports | Last seen |
|---|---|---|
| Handling ambiguity | 2 | Apr 2026 (report date) |
| Managing several deadlines | 1 | Sep 2026 |
| Decisions under uncertainty | 1 | Apr 2026 (report date) |
| Your most complex project | 1 | Apr 2026 (report date) |
| Accountability and owning failures | 1 | ~Dec 2025 |
| Conflict and cross-team work | 1 | ~Dec 2025 |
| Working with incomplete data | 1 | ~Dec 2025 |
| Leading under pressure | 1 | ~Dec 2025 |
11 detailed candidate timelines, loops, and screens
Newest first. G&L is Googleyness & Leadership and HC is the hiring committee.
01: Google L4 loop in team matching, with extra coding for an Applied AI team
When: report 2026-11-05 · Level: L4 · Outcome: in team matching, no match yet
- L4 onsite: not described.
- Extra coding (requested by an Applied AI team at team matching): two coding assessment rounds, format unknown, not held yet.
02: MLE two-day onsite with Decode String and a logging design
When: Sep 2026, a two-day onsite (round order unknown; part of the report is members-only) · Level: Mid Level / L4 (the site's label) · Where: US · Outcome: waiting at submission
- Coding ×2: the one named problem is Decode String with tree traversal (title only); practice version: decode a bit string with a code tree.
- Googleyness: managing several deadlines.
- AI/ML depth: some ML system design, though mostly the résumé and recent projects.
- Round type unclear: design a distributed logging system (title only); FastPrep files it under design a centralized log ingestion and search platform.
03: L4 loop for a similar role, described only by round type
When: report 2026-08-22 · Level: L4 · Outcome: not reported
- ML system design, coding, onsite coding rounds (count not given) and a hiring-manager team fit: no task details.
04: Google L4 SWE loop with two add-on coding rounds and an offer
When: add-on coding done shortly before 2026-06-12 · Level: L4 · Where: US · Outcome: offer at L4; joined, under a month in on 2026-09-17
- Domain: résumé-driven ML theory: hypotheticals, implementation details, and evaluation metrics with their definitions.
- Googleyness: behavioral questions.
- Onsite coding: a LeetCode medium using queues, with a follow-up.
- Onsite coding: a LeetCode medium DP variant, with a follow-up.
- Add-on coding: a medium/hard graph path-finding problem, with a follow-up.
- Add-on coding: a LeetCode hard DP problem, with a follow-up.
HC approved before team matching, and the one team-match call ended in a match.
05: ML Engineer loop with a Reviews-style design and two onsite coding rounds
When: report 2026-04-29 (two virtual rounds, then two onsite) · Level: L4 or L5 (unresolved) · Outcome: pending at posting
- ML system design (virtual): a Google Reviews-like system: the NLP pipeline, unsupervised clustering and summarization, an LLM vs a smaller model on cost, latency and quality, and a deep dive on BERT.
- Googleyness (virtual): decisions under uncertainty and ambiguity; the most complex project.
- Onsite coding (graphs + heaps): top-rated movie among transitively similar movies, ties broken lexicographically; follow-up: top K.
- Onsite coding (DP): count subsets whose sum is above half the total, example [1,3,4] → 3; FastPrep frames the same count as coalitions.
06: SDE3 / SWE III L4 AI/ML loop with two HC-requested add-ons, rejected
When: report 2026-03-30; the add-on rounds came before posting (original round order not given) · Level: L4 · Outcome: rejected after the loop, no reason given; six-month cooldown
- Googleyness, AI/ML and two coding rounds (virtual): no task details.
- Add-on coding (HC request): not described.
- Add-on Googleyness (HC request): not described.
Per the candidate, HC asked for the two add-ons because one interview left it short of signal.
07: L4 loop in the EU that closed during team matching
When: onsite late Mar to early Apr 2026 (inferred) · Level: L4 · Where: EU · Outcome: application closed by 2026-04-29, under four weeks into team matching; the closing email cited no feedback from teams
- Online coding, Googleyness, an AI domain round and two onsite coding rounds: no task details.
08: L4 / SWE III loop in EMEA still waiting on hiring-manager calls
When: end of Mar 2026 (round order not given) · Level: L4 · Where: EMEA; London, Munich, Paris and Zurich named · Outcome: still in team matching on 2026-07-17, with no hiring-manager calls
- Googleyness, three coding rounds and ML/AI depth: no task details.
09: Google L4 loop with an HC-requested technical round
When: phone screen to ML domain by 2025-10-10, leadership after that; HC-requested technical round 2026-02-17 · Level: L4 · Outcome: pending
- Phone screen, coding 1, coding 2, ML domain and leadership: no task details.
- HC-requested technical round (2026-02-17, after a team match): not described.
10: SWE III AI/ML loop with graph coding and an ML fundamentals round, rejected
When: loop ~Dec 2025 (inferred from a mid-Jan 2026 rejection call; every round 45 minutes) · Level: L4 · Outcome: rejected Jan 2026
- Phone screen: shortest path between two nodes; follow-ups: a changing graph, nodes blocked on the fly, a huge graph.
- Onsite coding: binary search on the answer, finding a min or max under a monotonic condition; follow-ups: why binary search, boundaries, large inputs, further optimization.
- Onsite coding: minimize the largest edge on a path, with threshold feasibility; follow-ups: why sum-Dijkstra fails, binary search plus BFS, a directed graph, path reconstruction, huge graphs.
- ML fundamentals & application: why-questions on supervised vs unsupervised learning, bias-variance, overfitting and regularization, metrics, feature engineering, clustering, entropy, hyperparameter search, black-box optimization, time series, and monitoring and retraining; less coding than talk. One-line scenarios: metrics for imbalanced data (related practice), deploying and monitoring a model (related practice), debugging a weak model, and designing an ML solution.
- G&L: accountability, ambiguity, conflict, cross-team work, incomplete data, owning failures, leadership under pressure.
11: SWE3 loop whose AIML domain round ran as system design
When: Googlyness 2025-10-13 · Level: L4 · Outcome: no decision communicated as of 2026-08-25
- Coding ×2: not described.
- AIML domain, run as system design: an image-classification app in a shared doc, covering architecture and models.
- Googlyness (2025-10-13): not described.
Patterns across the reports
- The hiring committee adds rounds. Reports 04, 06 and 09 each got extra rounds after the main loop: two add-on coding rounds (04), an add-on coding plus an add-on Googleyness round (06), and one more technical round after a team match (09).
- Team matching can stall. Report 07's application closed under four weeks into matching, citing no team feedback; report 08 was still waiting for hiring-manager calls almost four months after its loop; report 01 was still unmatched.
- Graph coding with many follow-ups. Reports 05 and 10 both had graph questions whose follow-ups changed the graph or the scale (top K in 05; blocked nodes, directed edges and huge graphs in 10).
Common questions
What rounds are in the Google SWE III ML-track interview?
Usually two or three coding rounds, an ML or AI domain round and Googleyness & Leadership, sometimes after a phone screen (04, 07, 08, 09, 10).
What is the Google ML domain round like?
It varied: résumé-driven ML theory (04), an ML system design such as a Reviews-like system (05), rapid why-questions on ML fundamentals (10), and an image-classification app run as system design (11).
Does the Google hiring committee ask for extra interviews?
In three reports it did (04, 06, 09). Report 06 was told one interview left the committee short of signal; 04 got an offer after its add-ons.
How long does Google L4 team matching take?
Report 04 matched on its first call; 07 was closed under four weeks in; 08 was still waiting with no hiring-manager calls almost four months after its loop.
What to practise
- Minimize the largest edge on a path and Shortest path between two nodes. Both come from report 10; graphs are the most repeated coding theme across the ML track (5 of 21).
- A Google Reviews-like system. ML design or applied-ML cases appear in 8 of the 21 ML-track reports.
Keep practicing
- Top-rated movie among transitively similar movies and Count subsets whose sum is above half the total: the two onsite coding questions from report 05.
- Decode a bit string with a code tree: the practice version of report 02's Decode String round.
- Google coding question bank: every Google coding problem on FastPrep (sign-in needed).
- Google ML design questions and Google system design questions (sign-in needed).
- Google interview guide and Google hiring insights: stages, practice problems and recent hiring signals.
- Experience feed: more interview reports from other companies and roles.
In this series: Part 1 (current) · Part 2 · Part 3
Sources: Anywhere useful interview info can be found.
Comments
0No comments yet — be the first to share what you know.