Tesla · Primly Community

Tesla onsite / final round: how it really goes, from someone who just finished it

sre_sol · 5 replies

Finished a 5-round virtual onsite for a senior ML role on the Autopilot AI team. Here's the honest breakdown.

Format: 5 rounds back-to-back over one day (with one 30-minute break in the middle). Fully virtual, Zoom. Each round was exactly 60 minutes.

My five rounds: Coding (algorithms, 1 problem) Coding (domain-specific, involved matrix operations) ML system design (designing a perception pipeline component) Behavioral Hiring manager conversation (more strategic, less testing)

Other roles might have different mixes. A product SWE friend who did their onsite around the same time had 2 coding, 1 system design, 1 behavioral, 1 HM. No second domain-specific coding. So the exact mix depends on the team.

Pacing: The day is long. By round 4 I was tired. This matters. Bring water. Take the break seriously. I ate something actual during my 30 minutes and it helped.

Interviewers: All were engineers or engineering leads on the actual team, not HR or generalist interviewers. One of my coding interviewers had been at Tesla for 4 years on Autopilot software. That means the questions are relevant to real problems, but it also means they have opinions. When I proposed a solution, they pushed back with "we actually tried that approach and hit X problem." That's both challenging and kind of great because you get real signal on how they think.

What matters most: In my experience the behavioral and HM rounds carry a lot of weight in the final debrief, not just the coding scores. Tesla seems to weight culture and drive pretty heavily. A friend who aced the coding rounds but seemed passive in the behavioral didn't get an offer. Someone I know who flubbed one coding problem but was sharp and direct in behavioral and HM did get through.

Timeline after onsite: They told me 5-7 business days. Actual response came in 9 business days. One extension email in between. Don't read too much into the delay.

5 replies

qa_quinn

"Interviewers had opinions" is such a specific Tesla thing. At my old FAANG shop the interviewers were explicitly told not to share their own approach during the problem. Tesla being more conversational is double-edged -- it's more human but also means there's more variance depending on who you get. Did you feel like different interviewers were evaluating different things or was there obvious calibration?

ml_mike

Seemed calibrated on the coding rounds (rubric-y, they took consistent notes). Less calibrated on behavioral: one interviewer asked deeply about a specific past project for 40 of 60 minutes, another ran through 5 different situations. Hard to know if that's a problem or just interviewer style. I came out not sure how the behavioral would land. It did land fine so maybe varied depth is okay.

hardware_hugo

How ML-specific was the domain coding round? Like are we talking LeetCode with matrices or actual ML-adjacent stuff like implementing a loss function or building a simple inference pipeline?

ml_mike

More the latter. I was asked to implement a version of non-maximum suppression (common in object detection postprocessing) and reason through edge cases. Not a leetcode graph problem. Very domain-specific. For Autopilot ML roles I'd brush up on computer vision fundamentals and be comfortable implementing standard CV algorithms from scratch.

finance_faye

The behavioral carrying weight in the debrief tracks with what I've heard from people inside Tesla's hiring process. They have a culture of speed and directness and the debrief can get killed by someone who felt the candidate was slow or passive. The behavioral round is not a checkbox. Take it as seriously as the coding.