Every entry below is a crawlable link to a unique public task page. This directory complements the interactive filters and keeps the complete catalog connected through standard pagination.
Tech Screening · Hard · 3 min
Which data checks belong before training starts, and which failures should block a production model release?
View taskTech Screening · Easy · 3 min
A categorical feature gains thousands of unseen values overnight. How would your pipeline surface ownership and impact?
View taskTech Screening · Medium · 3 min
Fraud labels arrive thirty days late. How would you evaluate current model quality before the final labels exist?
View taskTech Screening · Hard · 3 min
How would delayed and selectively observed labels affect retraining, monitoring, and experiment decisions?
View taskTech Screening · Easy · 3 min
A larger model improves offline quality slightly but doubles serving cost. How would you make the launch decision?
View taskTech Screening · Medium · 3 min
How would you compare two models when aggregate quality, tail latency, calibration, and cohort performance disagree?
View taskTech Screening · Hard · 3 min
How would you estimate whether an online experiment can detect the smallest model improvement worth shipping?
View taskTech Screening · Easy · 3 min
An experiment is neutral after three days. What would you check before calling the treatment ineffective?
View taskTech Screening · Medium · 3 min
A ranking experiment lifts clicks but increases hides and latency. How would you decide whether to continue?
View taskTech Screening · Hard · 3 min
Which primary, guardrail, and diagnostic metrics would you define before testing a recommendation model?
View taskTech Screening · Easy · 3 min
When would NDCG be more useful than precision at k, and what user behavior can neither metric capture?
View taskTech Screening · Medium · 3 min
Offline ranking metrics improve while session satisfaction falls. How would you reconcile the two signals?
View taskTech Screening · Hard · 3 min
Why do large recommendation systems separate candidate retrieval from ranking, and where can quality be lost?
View taskTech Screening · Easy · 3 min
Candidate recall drops after an embedding update. How would you determine whether retrieval or ranking caused the regression?
View taskTech Screening · Medium · 3 min
How would you serve useful recommendations for a new user with no behavioral history?
View taskTech Screening · Hard · 3 min
A collaborative model performs well overall but poorly for new items. What fallback and evaluation would you add?
View taskTech Screening · Easy · 3 min
How would you evaluate whether a new embedding space is useful beyond looking at nearest-neighbor examples?
View taskTech Screening · Medium · 3 min
Embedding norms shift after retraining and retrieval quality falls. What evidence would you inspect before rolling back?
View taskTech Screening · Hard · 3 min
Aggregate accuracy is stable, but one protected cohort regresses. How would you investigate and frame the launch decision?
View taskTech Screening · Easy · 3 min
How would you choose fairness slices and metrics without turning the review into a checkbox exercise?
View taskTech Screening · Medium · 3 min
When is a feature-attribution explanation useful, and when can it create false confidence?
View taskTech Screening · Hard · 3 min
A regulated decision requires reason codes. How would you validate that explanations are stable and faithful enough?
View taskTech Screening · Easy · 3 min
A useful training feature contains sensitive user data. How would you challenge its necessity and reduce exposure?
View taskTech Screening · Medium · 3 min
How would deletion requests propagate through datasets, features, trained models, and cached predictions?
View taskTech Screening · Hard · 3 min
What metadata must a model registry preserve so an on-call engineer can reproduce and roll back a release?
View taskTech Screening · Easy · 3 min
A production endpoint serves an unexpected model artifact. How would you trace code, data, features, and deployment lineage?
View taskTech Screening · Medium · 3 min
When would you use shadow traffic instead of a canary for a new model, and what can shadowing not prove?
View taskTech Screening · Hard · 3 min
Design a staged model rollout with explicit promotion and rollback signals.
View taskTech Screening · Easy · 3 min
What must be versioned together for a model rollback to be genuinely safe?
View taskTech Screening · Medium · 3 min
You roll back the model, but quality remains poor. Which coupled dependencies would you investigate next?
View taskTech Screening · Hard · 3 min
Which production signals would you monitor when ground-truth labels arrive weeks later?
View taskTech Screening · Easy · 3 min
A model endpoint is healthy, but business quality is deteriorating. How would your monitoring make that visible?
View taskTech Screening · Medium · 3 min
Model p99 latency rises while average latency is stable. How would you find the bottleneck?
View taskTech Screening · Hard · 3 min
Quality improves with a larger model, but it misses the serving budget. Which optimization options would you compare?
View taskTech Screening · Easy · 3 min
How does dynamic batching trade throughput against latency, and which workload signals control the decision?
View taskTech Screening · Medium · 3 min
GPU utilization is low while request latency is high. How would you test whether batching or queueing is misconfigured?
View taskTech Screening · Hard · 3 min
When is caching model predictions safe, and what should be part of the cache key?
View taskTech Screening · Easy · 3 min
A prediction cache improves latency but serves stale personalization. How would you set invalidation and freshness rules?
View taskTech Screening · Medium · 3 min
What does reproducible training require beyond setting a random seed?
View taskTech Screening · Hard · 3 min
A past model cannot be rebuilt after an incident. Which data, code, environment, and configuration artifacts were missing?
View taskTech Screening · Easy · 3 min
How would you design a hyperparameter search without overfitting to the validation set?
View taskTech Screening · Medium · 3 min
A broad search finds a tiny metric gain at large compute cost. How would you decide whether the result is credible?
View taskTech Screening · Hard · 3 min
Training quality improves while validation quality plateaus. What evidence would distinguish overfitting from a data bug?
View taskTech Screening · Easy · 3 min
A high-capacity model wins one split but loses across time and cohorts. How would you redesign the evaluation?
View taskTech Screening · Medium · 3 min
Inputs look stable, but the relationship between features and outcomes changes. How would you detect concept drift?
View taskTech Screening · Hard · 3 min
Seasonal behavior makes a model appear degraded. How would you separate expected cycles from a real decision-boundary change?
View taskTech Screening · Easy · 3 min
How can a recommender create a feedback loop that makes its own offline metrics look better?
View taskTech Screening · Medium · 3 min
Exposure determines which labels you observe. How would you reduce bias when training the next ranking model?
View task