← Back to Directory

TikTok interview notes, page 24 of 28

Large Enterprises

Most of these TikTok write-ups come from Data Scientists (111), then Software Engineers (102) and Product Managers (35). The most reported rounds are the technical phone screen (167), the product sense round (33), the system design round (25) and the behavioral / leadership round (21).

279 interview notes · updated Sep 2026

TikTok·Machine Learning Engineer·Technical Phone Screen

Jun 2026
TikTok ML engineer interview that went pretty deep on Transformer internals. They wanted pseudocode, not just hand-waving, which I wasn't fully prepared for. The concept questions at the end felt almost like a relief after writing out attention by hand.
  • Write pseudocode for scaled dot-product self-attention for a single head, including tensor shapes, Q/K/V projections, scaling, softmax, and optionally an attention mask.
  • Write pseudocode for multi-class cross-entropy loss over a batch, handling both class index labels and one-hot labels, returning a scalar mean loss.
  • What does the position-wise feed-forward network in a Transformer block actually do, and why is it necessary given that attention already exists?
  • Why do we divide attention scores by the square root of the head dimension before applying softmax?

“This is where I spent most of my mental energy.”

View Post

TikTok·Data Scientist·Technical Phone Screen

Jun 2026
TikTok data scientist interview with a heavy SQL focus. Three questions, all interconnected around the same dataset, which I wasn't expecting. Felt more like a take-home problem crammed into a live session.
  • Write an SQL query to compute a 7-day rolling average of daily unique viewers for every post, ordered by view date.
  • Given a search term like 'Apple', write a query to return all post IDs where the hashtags column contains that term, case-insensitively.
  • What is the logical execution order of SQL clauses, and why does understanding that order matter for debugging or optimization?

“The window function part I got right, COUNT(DISTINCT user_id) over a RANGE between 6 preceding and current row.”

View Post

TikTok·Machine Learning Engineer·Technical Phone Screen

Jun 2026
Conceptual ML round at TikTok for an MLE role. Three questions, all theory-heavy, no coding. Felt more like a grad school oral exam than a typical tech screen.
  • Walk me through the main components of a Transformer block and explain what each one does.
  • What are the key architectural and pretraining differences between GPT and BERT, and when would you choose one over the other?
  • Define precision and recall, then explain how adjusting a classification threshold shifts the tradeoff between them. Give examples of when you'd prioritize each.

“Knew this one cold.”

View Post

TikTok·Machine Learning Engineer·Onsite - System Design / Architecture

Jun 2026
ML engineer interview at TikTok with a product design question focused on the recommendation system behind the For You page. Pretty open-ended, which I wasn't fully prepared for coming from a more traditional ML background.
  • How would you design the 'For You' recommendation feed on TikTok?

“I went straight to the ML pipeline: candidate generation, ranking, re-ranking for diversity.”

View Post

TikTok·Machine Learning Engineer·Technical Phone Screen

Jun 2026
TikTok ML engineer round that was basically a deep-dive into one hard geometry/algorithm problem. They wanted a full working solution plus complexity analysis and edge case handling, not just a sketch. Felt more like a systems-adjacent coding round than a pure ML interview.
  • Given n axis-aligned rectangular buildings each described by a left x-coordinate, right x-coordinate, and height, compute the outer silhouette (skyline) of all buildings combined. Return the list of critical points where the height changes, sorted by x, with no two consecutive points at the same height, and the last point dropping to zero. Aim for O(n log n) time. Implement an efficient approach such as a sweep line with a max-heap or divide-and-conquer, and explain your design. Also specify how you handle ties when multiple building edges share the same x-coordinate: entering edges before leaving, among entering edges process taller buildings first, among leaving edges process shorter buildings first. Discuss correctness, complexity, and edge cases including identical buildings, fully nested intervals, buildings that share a border, and very large coordinate values.

“This is the classic skyline problem but they pushed way further than I expected.”

View Post

TikTok·Data Scientist·Technical Phone Screen

Jun 2026
TikTok data scientist technical screen, heavy on applied ML evaluation theory. The whole thing was basically one long question about a classifier operating on an imbalanced dataset, broken into four parts. Left feeling like I'd done okay on parts one and two but probably fumbled the derivation at the end.
  • Given a calibrated classifier with 1% positive class prevalence and a held-out set of 10,000 examples, compute Precision, Recall, and F1 at two different thresholds (0.50 and 0.20) using the provided confusion matrix counts.
  • Using a cost matrix where a false positive costs 1 and a false negative costs 20, calculate the expected total cost at each threshold and determine which threshold is cheaper.
  • Why can ROC-AUC be misleading on a heavily imbalanced dataset, and why is PR-AUC a better choice? Use the numbers from the confusion matrices to build a quick intuition.
  • Derive the optimal probability threshold t* for a cost-sensitive decision using a perfectly calibrated model, expressed in terms of the false positive cost, false negative cost, and the class prior.

“This part was fine.”

View Post

TikTok·Data Scientist·Technical Phone Screen

Jun 2026
A deep technical screen for a DS role at TikTok that was basically one giant causal inference design problem. The question covered so much ground it felt less like an interview and more like a take-home compressed into 45 minutes.
  • Design an incrementality test from an advertiser's perspective to measure whether TikTok ads actually drive lift in conversions. Walk through your randomization unit and why, how you'd construct the holdout group, your primary KPIs and their formulas, sample size and duration math, how you'd handle cross-channel spillovers and auction effects, what guardrails and falsification checks you'd put in, and how you'd report heterogeneous lift without p-hacking.

“This was basically seven questions welded into one.”

View Post

TikTok·Data Scientist·Technical Phone Screen

Jun 2026
TikTok data scientist interview that was basically one giant SQL question with a lot of moving parts. The schema looked simple but the actual requirements kept stacking up and I spent more time than I'd like to admit untangling the timezone logic.
  • Write a single PostgreSQL query to produce a 7-day regional dashboard (US vs Asia) with columns for local date, DAU, buyers, revenue, ARPPU, view-to-purchase conversion rate, new vs returning DAU, and a flag for days where revenue drops more than 20% below the 7-day median for that region. Use per-user timezone offsets to define 'local date', treat 2025-09-01 as today, and ensure the query is idempotent against duplicate events sharing the same user, timestamp, event type, and session.

“This took me way longer to structure than I expected.”

View Post

TikTok·Data Scientist·Technical Phone Screen

Jun 2026
TikTok DS interview with a deep-dive experiment design question around changing the default video length. One question, very open-ended, and they clearly wanted you to get into the weeds on instrumentation and rollout, not just wave at 'run an A/B test'.
  • Product wants to change the default video length from 15 seconds to 60 seconds. Walk through how you'd design this experiment end to end: randomization unit and contamination risks across creators and viewers, primary metrics and guardrails, traffic split and ramp schedule, how you'd handle novelty effects, SUTVA checks, and what a post-experiment rollout looks like.

“This one is genuinely hard because TikTok is a two-sided platform and your randomization choice breaks everything downstream.”

View Post

TikTok·Data Scientist·Technical Phone Screen

Jun 2026
TikTok data scientist interview, SQL-heavy technical screen. One big multi-part query problem that took up basically the whole session. Not a vibe check at all, just pure SQL under pressure.
  • Given a sellers table and an events table with order and complaint records, write a single Postgres query (no procedural code) to compute a 7-day rolling complaint-to-order ratio per seller per day. The window for each date d is [d-6, d] inclusive, events are bucketed by date, and if the order count in the window is zero the ratio should be NULL. Also produce a second result row aggregating across all sellers for a specific reference date, with seller_id shown as 'ALL'.

“This wrecked me a little.”

View Post