just wrapped the Fidelity coding rounds last month. sharing what I saw because a lot of the older forum posts describe a process that doesn't quite match what's happening now.
the OA (online assessment)
two problems on HackerRank. 90-minute window. you schedule it yourself within a few days of getting the link. I got: a graph traversal problem (find shortest path in a weighted graph, standard BFS/Dijkstra territory) a string parsing problem that was more implementation-heavy than algorithmic
difficulty: call it leetcode medium on both. not hard. if you've done 50-80 medium problems recently you should be fine. the OA is more of a bar-check than a differentiator. the scoring didn't feel like it cared about edge cases as much as "did you get the correct output on the visible test cases."
there are also a small number of multiple choice questions about Java and some data structures basics. these were trivial but don't skip them, they're timed per question.
the technical phone screen coding portion
this is different. one problem in CoderPad (live, shared). the problem I got was also medium-difficulty but the expectation here is clearly different: they want to see how you think, not just whether you can output code. I got stuck briefly on edge case handling and talked through it out loud and the interviewer helped me past it. that's fine and expected. don't freeze and go silent.
the onsite coding round
still one algorithmic problem but now with a design/extension layer. so you solve the core coding problem, then they ask you to extend it: "now what if the input could be millions of rows?" or "what if you needed to add X feature?" this tests whether you can think beyond the immediate solution.
a few notes on language choice: I used Python. nobody batted an eye. I've heard Java is common internally and might land better for some interviewers but I have no hard data on that. ask your recruiter if it matters for the specific team.
what's changed: at least one person I talked to said the OA used to be harder (3 problems, more algorithmic) and the behavioral section used to be a separate call. it looks like they consolidated things. the current format feels more calibrated: OA weeds out people who haven't coded in years, the phone screen and onsite do the actual signal-gathering.
overall difficulty: solidly medium. not a grind-500-leet-code process. but don't go in cold.