Applied ML Scientist · Full Journey · Multiple Types

Microsoft — Applied ML Scientist ✅ Passed

Level: Staff-Level

Round: Full Journey · Type: Multiple Types · Difficulty: 8/10 · Duration: 480 min · Interviewer: Unfriendly

Topics: Machine Learning, Large Language Models, System Design, Behavioral Questions, Random Processes, Vector Semantic Search, Preference Alignment, Retrieval Augmented Generation (RAG)

Location: San Francisco Bay Area

Interview date: 2025-11-15

Summary

Interview Rounds Overview

• Round 1: Principal AS Phone Screen (Dynamics)
• Round 2: Principal Phone Screen (Dynamics)
• Round 3: Loop (Dynamics)
• Round 4: Loop (Word)
• Round 5: Loop (Outlook)
• Round 6: Loop (Copilot Tuning)
• Round 7: Loop (Copilot Shared Services)
• Round 8: Loop (Unknown Group)

Details

I received a phone screen from a Principal AS from the Dynamics group in late August 2025. I felt it went well, but I didn't hear back. Then, I interviewed with an Indian manager, also a Principal, from another Dynamics group. The process went smoothly, but again, I received no response. Around September 20, the first Dynamics group suddenly scheduled a loop interview. This was my first Microsoft loop.

The loop was extremely difficult, involving deriving stochastic processes and very challenging LLM questions. I managed to handle them and felt good about my performance. I thought these questions were beyond what is normally asked in an interview. My overall impression of this group was good.

However, I didn't hear back for two weeks after the loop. In the meantime, I interviewed with the Word and Outlook loops. In the third week, after the Outlook interview, the Dynamics recruiter contacted me and said that the team approved me and that I passed the interview, but they couldn't offer a Level 65 position, only a Level 64, and needed to redo the phone screen and loop. After I confirmed my willingness to redo the loop, they blocked the inclined Outlook group from proceeding. In the end, the Outlook group disappeared.

The new loop was basically a formality: the interviewers I had already met simply greeted me and left. Only the HM round became a reverse behavioral question session. Then, we talked about compensation. When I tested the waters, the recruiter vaguely mentioned "xxx," which I said I could accept. In the first round of formal compensation talks, they only offered 60% of the previous number (why did they say "xxx" before?). They said the compensation team required offers from other companies to negotiate. After I provided other offers, it increased by 30%, mainly in RSU, which increased by 250%, but it was still 20% lower than "xxx." They told me that there would be a sign-on bonus in the first year, but no bonus, and that the second year would be the highest TC year. Near the end of the consideration period, the recruiter scheduled another chat, first saying that it was impossible to add any more money. However, the details were completely changed:

1. All the previous numbers calculated did not include the bonus, but I could (again) have a bonus in the first year??
2. The second year was not the highest. Refreshers were also considered as a bonus, so the fourth year was the highest TC year, because of the cumulative refreshers from the previous three years.

While negotiating with Dynamics, I interviewed with three more groups. The last two groups didn't have phone screens before the loop. The HMs were all SDEs. One group had SDEs asking standard interview questions and random questions throughout the process. Another group had an AS from another group interview me, and no one from the group itself showed up. Honestly, I didn't understand it at all. Now, I will briefly summarize the seven group loops I interviewed with:

1. Dynamics: I had to derive random process formulas in Teams during one round.
2. Dynamics: Downgraded and then put on hold.
3. Word: One round focused on standard interview questions, and another was based on Word-related case studies, mainly asking how to calibrate LLM-generated formats that do not conform to Word formats.
4. Outlook: One round involved ML coding, one round had the HM asking about my resume, one round had two people listen to my experience presentation, and one round involved Outlook case studies, mainly discussing how to de-bias email display order and personalize email priority and summaries. I passed these four rounds.
5. Copilot Tuning: About half of the people in this group were Brazilian. One coding round was very challenging, one round was on LLM case studies, one round discussed vector semantic search optimization, and one skip behavioral question round.
6. Copilot Shared Services: This loop only had three rounds and no phone screen. After the interview, I was ghosted. One round mainly discussed the pros and cons of various preference alignment methods, one round involved standard ML questions, and one round discussed my resume. However, the SD at the end was a bit tricky, asking me to guess the user's intent based on their typing. This was not ordinary search completion text but rather intent understanding. None of the interviewers were from this group, as the group consisted entirely of SDEs.
7. This round had the worst experience. It was all SDEs interviewing me because the entire group was SDEs. It's not meant to disparage SDEs, but the questions were completely irrelevant. They asked about my resume in a way that missed the point, and the questions had a strong backend flavor, followed by behavioral questions. I forgot what the product of this group was. The technology was RAG, and I was suddenly scheduled for a loop without a phone screen.