Interview Experience | Data Infra | π₯π₯π₯ | Figure
2026 (Oct-Nov) Helix AI Engineer, Data Infrastructure @ Figure β San Jose, CA β Final round β 3 rounds | Hard
Round 1 β Pair Programming
Two engineers, Python, shared editor.
The problem: one recording has two streams, camera frame timestamps frames and joint state timestamps joints, both sorted integers in ms.
Part 1: for each frame, return the index of the nearest joint sample; on a tie take the earlier one.
frames = [10, 25]
joints = [8, 12, 30]
-> [0, 2] 10 is equally close to 8 and 12, take 8; 25 is closest to 30
Two pointers, joint pointer only moves forward. On his case the last frame ran the pointer off the end once; added a bounds check.
Part 2: add a tolerance tol; |dt| > tol doesn't count. Each joint sample can go to at most one frame, training wants one-to-one. Max number of matched frames?
frames = [6, 8], joints = [4, 7], tol = 2
6 -> 4, 8 -> 7, answer 2
I kept going from part 1: each frame takes the nearest unused sample. He gave exactly this case: 6 is closer to 7 and takes it, 8 has nothing left, answer 1.
Switched to: frames in order, each takes the earliest sample it can still match. Skip joints < f - tol, and if the first remaining one is <= f + tol, match it. Earlier samples are even less useful to later frames, so leave the later ones for them. He asked why it's correct; gave an exchange argument: if an optimal solution matches a frame to a later sample, swapping to the earliest feasible one doesn't hurt anything after. Came out a bit tangled, he said OK.
Part 3: among maximum matchings, minimize total |dt|. Said optimal pairs don't cross, so DP, dp[i][k] over first i frames and first k joints, O(nm). He asked for faster, both streams are long. Thought for a while, didn't get it, time.
Round 2 β Code Review
One engineer, hands you a PR, you review and leave comments. No running it, just read and call out problems.
The problem: the PR adds two functions that prepare training data.
make_splits(clips, val_ratio): shuffle clips, cut into train / val by ratiobuild_sampler(clips): weight each clip1 / count[clip.task]so every task gets sampled about equally
A clip is a fixed-length slice of an episode; one episode yields many clips:
episode A (folding laundry) -> a1 ... a40
episode B (wiping a table) -> b1, b2, b3
First I flagged duration == 0, said you can't compare floats like that. He said duration is integer nanoseconds, fine. Lost a few minutes.
The real problem is the split. Random by clip means a1 in train and a2 in val, a few seconds apart and nearly identical, so val scores look too good. At minimum split by episode. He pushed: is episode enough? Episodes from the same robot, same kitchen, same day look alike too. Said group by (robot, scene, date), each group goes entirely to one side.
The weights: 1 / count[task] counts clips, so tasks are balanced, but inside a task long episodes like A still dominate. Asked whether he wants tasks or episodes balanced; he said both. Changed to three levels: by task, then by episode, then uniform within the episode.
Also clips.sort(...) sorts in place, and clips is the same list from a cache, so it changes data someone else is holding. Only saw that after his hint, "look at where the argument comes from".
Last: after a label fix, can old models' scores on the old val be compared with new ones. Said freeze val as a versioned manifest. How label versions map to manifests, didn't finish.
Round 3 β System Design
Design robot data offload: each robot records far more per day than can be uploaded, and it only has network while charging.
Requirements:
- data training needs should come up fast
- researchers can later ask for raw data "around some event"
- when the robot's disk fills it has to delete, but nothing unconfirmed in the cloud may be deleted
First drew an upload queue on the robot, FIFO by record time into object storage, with an episode table in the cloud. He said bandwidth only covers a small part, so FIFO is basically random dropping.
Changed to two tiers. Metadata first: times, task, event markers, low-res thumbnails, small enough to finish every charge. Raw data stays on the robot and goes up by priority, from two sources: events the robot flags itself (fall, e-stop, failed grasp) pin a window around them to the top; researchers look at metadata in the cloud and send a pull request for a segment.
I said delete locally once uploaded. He asked what "uploaded" means; I said the API returns 200. He said a 200 on the last part of a multipart upload doesn't mean the object is complete. Changed to: cloud completes, verifies checksum, writes the episode table, sends a confirmation; the robot marks data deletable only after that.
Deletion order: confirmed data first, then unrequested low-priority data. Anything pinned by a pull request or an event stays.
He asked: a researcher wants a segment from three days ago and the robot already deleted it. Said that's why metadata goes first; metadata carries "still on robot", flipped when deleted, so researchers can only request what's there.
Last: every team pins stuff, pinned data exceeds the disk. Said per-team quotas, over quota release the oldest first. How to set quotas and who decides, didn't finish.
For finding roles and applying, LandOffer.ai aggregates job postings and auto-fills applications for free.