Top 30 Advanced Data Science & Analytics Interview Questions and Answers

30 questions. Pick an answer, or open “Show the answer” to read it.

Answers are graded in your browser. Nothing is saved, and no XP or streak is earned here. The app keeps score.

1. Question 1 of 30

   Why is the correlation between profile completion and retention weak evidence that forcing completion will improve retention?

   • aCorrelation coefficients cannot apply to user behavior data
   • bRetention can never be measured accurately for new users
   • cThe sample of completed profiles is always too small to trust
   • dEngaged users self-select into completing profiles, confounding the relationship
   Show the answer

   Answer: d · Engaged users self-select into completing profiles, confounding the relationship

   Motivated users both complete profiles and retain, so engagement is a confounder making completion a marker rather than a proven cause. A randomized experiment is needed; the other options misstate the actual problem.

   Read the full bite: Does forcing profile completion cause retention?

2. Question 2 of 30

   What is the most important addition to a revenue metric when evaluating a change that may harm long-term satisfaction?

   • aMeasuring revenue more precisely with a larger sample
   • bLong-horizon guardrail metrics like retention plus a sufficiently long experiment
   • cShipping quickly to capture the revenue before users notice
   • dRelying on stakeholder intuition about user happiness
   Show the answer

   Answer: b · Long-horizon guardrail metrics like retention plus a sufficiently long experiment

   Long-term harm surfaces as delayed churn, so guardrail metrics measured over a long-enough window are essential to weigh against the immediate revenue lift. More precise short-term revenue alone still misses the delayed retention cost.

   Read the full bite: Framing ad-load tradeoffs: revenue versus retention

3. Question 3 of 30

   When validating a composite burnout proxy, why test whether it predicts voluntary attrition six months later?

   • aIt confirms the survey instrument has higher reliability than behavioral telemetry
   • bIt proves burnout causally drives attrition rather than merely correlating with it
   • cIt establishes predictive validity by showing the proxy correlates with a meaningful future outcome
   • dIt satisfies convergent validity by aligning the proxy with self-reported measures
   Show the answer

   Answer: c · It establishes predictive validity by showing the proxy correlates with a meaningful future outcome

   Testing future attrition establishes predictive validity, confirming the composite index captures a construct with real downstream consequences. The most tempting distractor confuses prediction with causation: a proxy that predicts attrition does not prove burnout causes it, since unobserved confounders may drive both.

   Read the full bite: How would you build and validate a proxy target for employee burnout?

4. Question 4 of 30

   In PCA on mean-centered data, what does the eigenvalue associated with a principal component directly represent?

   • aThe number of samples lying on that axis
   • bThe correlation between the original two features
   • cThe classification accuracy gained by keeping that component
   • dThe variance of the data projected onto that component's direction
   Show the answer

   Answer: d · The variance of the data projected onto that component's direction

   Each eigenvalue equals the variance captured along its eigenvector, which is why components are ranked by eigenvalue. Accuracy is a downstream model metric, not what the eigenvalue measures.

   Read the full bite: Eigenvalues, eigenvectors, and their role in PCA

5. Question 5 of 30

   In a two-proportion conversion test, which change most directly explains why detecting a 2% relative lift requires roughly 390k users per variant instead of 63k for a 5% relative lift?

   • aThe absolute MDE shrinks from 0.5 to 0.2 percentage points, and sample size scales roughly with the inverse square of the absolute effect size.
   • bLowering the relative lift from 5% to 2% reduces statistical power from 80% to approximately 20% unless sample size is inflated.
   • cThe standard error of a proportion depends on p(1-p), which drops nonlinearly when the treatment rate approaches the baseline rate.
   • dThe baseline conversion rate of 10% creates a ceiling effect that exponentially increases variance as the relative lift decreases.
   Show the answer

   Answer: a · The absolute MDE shrinks from 0.5 to 0.2 percentage points, and sample size scales roughly with the inverse square of the absolute effect size.

   The correct answer identifies that the absolute MDE falls from 0.5 to 0.2 percentage points, and because N is roughly proportional to the inverse square of the absolute MDE, required sample size balloons by about sixfold. Distractor D is tempting because it quotes the true p(1-p) variance formula from the card, but in this range variance barely changes and cannot explain the explosion in sample size.

   Read the full bite: How do you determine sample size for a conversion lift experiment?

6. Question 6 of 30

   Under what condition does MAP estimation give essentially the same result as MLE?

   • aWhen the prior is uniform or the dataset is very large
   • bWhen the dataset is very small
   • cWhen the likelihood is non-convex
   • dWhen the prior is very strong and informative
   Show the answer

   Answer: a · When the prior is uniform or the dataset is very large

   A uniform prior makes the posterior proportional to the likelihood, and with large data the likelihood overwhelms any prior, so MAP converges to MLE. A strong prior on small data does the opposite.

   Read the full bite: MLE versus MAP estimation and the role of priors

7. Question 7 of 30

   When aggregating a 50GB CSV on a 16GB machine, which strategy keeps peak memory usage proportional to a small fragment rather than the entire file?

   • aUse read_csv with usecols and dtype downcasting, then load everything into one DataFrame
   • bSample 1000 rows with nrows to inspect dtypes, then process the full file with default parameters
   • cIterate with read_csv(chunksize=...), aggregating each fragment and discarding it before reading the next
   • dLoad the full file with optimized dtypes but rely on OS swap for overflow
   Show the answer

   Answer: c · Iterate with read_csv(chunksize=...), aggregating each fragment and discarding it before reading the next

   Streaming with chunksize processes only one fragment at a time, keeping memory bounded by that fragment instead of the full 50GB. Option A is tempting because filtering columns and downcasting dtypes are valid optimizations, but materializing the entire file in a single DataFrame still exhausts RAM.

   Read the full bite: Process a 50GB CSV with only 16GB RAM

8. Question 8 of 30

   You convert a 10-million-row DataFrame's string column with 6 million unique values to category. What is the likely effect on memory usage?

   • aMemory drops significantly because category stores integer codes instead of full strings.
   • bMemory increases because the category codes plus unique values list outweigh a simple object array.
   • cThe conversion raises a ValueError because category dtype cannot represent more than fifty percent unique values.
   • dMemory stays roughly the same as object dtype since code overhead balances string storage.
   Show the answer

   Answer: b · Memory increases because the category codes plus unique values list outweigh a simple object array.

   When cardinality exceeds roughly fifty percent of row count, the storage cost of category codes plus the unique values list exceeds that of a plain object array. Option A is a common misconception that category dtype always reduces memory, but it is only beneficial for low-cardinality columns.

   Read the full bite: How do you analyze and reduce large pandas DataFrame memory usage?

9. Question 9 of 30

   Which pattern best synchronizes a large OLTP table to a warehouse while correctly handling hard deletes, out-of-order updates, and exactly-once recovery?

   • aAdd a deleted_at column to every source table and filter on updated_at, converting hard deletes to updates before extraction
   • bDual write from the application service to both the OLTP database and the warehouse, retrying warehouse failures asynchronously
   • cPoll the source using an updated_at high watermark and run a daily full extract to reconcile deletes and late arrivals
   • dCapture database change events from the transaction log, stage them in an open-table format, and merge incremental files into the warehouse with checkpointed offsets
   Show the answer

   Answer: d · Capture database change events from the transaction log, stage them in an open-table format, and merge incremental files into the warehouse with checkpointed offsets

   CDC from the transaction log captures every mutating event including deletes and preserves ordering, while staging in an open-table format enables incremental merge into the warehouse with checkpointed offsets for exactly-once recovery. Dual writes risk inconsistency across independent transactions, and simple timestamp watermarking misses hard deletes and struggles with out-of-order rows unless paired with costly full extracts.

   Read the full bite: Design an incremental load pipeline from a transactional DB to a warehouse

10. Question 10 of 30

    When targeting a site protected by advanced bot management that fingerprints TLS, browser runtime, and behavior, which strategy is most effective?

    • aRun vanilla headless Chrome and add uniform random sleeps between actions
    • bMatch JA3/TLS signatures, patch automation leaks, and use realistic mouse trajectories
    • cWarm sessions by browsing entry pages and rely on CAPTCHA-solving services
    • dRotate residential proxies and randomize user-agent strings per request
    Show the answer

    Answer: b · Match JA3/TLS signatures, patch automation leaks, and use realistic mouse trajectories

    Modern anti-bot stacks fingerprint TLS handshakes, browser automation leaks, and interaction entropy, so evasion requires addressing all three layers simultaneously. Proxies and user-agents alone fail at the transport layer, vanilla headless Chrome exposes runtime signals, and relying on CAPTCHA solving ignores the root detection vectors.

    Read the full bite: How would you evade an advanced anti-bot system while scraping?

11. Question 11 of 30

    Which design decision most directly prevents a surge from thousands of IoT devices from overwhelming downstream stream processors?

    • aIncreasing topic partition count to match the number of connected devices
    • bWriting raw telemetry directly to cold storage instead of the streaming buffer
    • cConfiguring Kafka with three replicas per partition and producer acks set to all
    • dPlacing local rate limiting and batching at the edge gateway before the Kafka backplane
    Show the answer

    Answer: d · Placing local rate limiting and batching at the edge gateway before the Kafka backplane

    The edge gateway isolates the core pipeline by applying local rate limiting and batching before Kafka. While three replicas with acks=all ensures durability against rack failures, it does not throttle an incoming flood from thousands of devices.

    Read the full bite: Design a scalable, fault-tolerant real-time IoT data ingestion system

12. Question 12 of 30

    In a production time-series forecasting pipeline using rolling-origin validation, which approach correctly prevents future leakage?

    • aUse backward-looking windows ending at t minus one and fit scalers exclusively on each training fold before transforming the matching validation fold
    • bApply standard k-fold cross-validation after generating lag features to maximize the number of performance estimates
    • cReuse the scaler parameters from the first training window for all later folds to ensure consistent feature distributions
    • dCompute centered rolling means and fit preprocessing globally across the entire dataset before splitting
    Show the answer

    Answer: a · Use backward-looking windows ending at t minus one and fit scalers exclusively on each training fold before transforming the matching validation fold

    Backward-looking windows ending at t minus one ensure no future data enters features, and fitting preprocessing per training fold stops global statistics from leaking into validation. Option D is tempting because global scaling is standard in non-temporal ML, yet it embeds future information into every historical row before any split occurs.

    Read the full bite: How do you prevent future leakage in time-series preprocessing?

13. Question 13 of 30

    While designing a two-pass standardizer for a 500GB out-of-core dataset, your colleague proposes computing each partition's mean and then averaging those means with equal weights. What is the fundamental flaw?

    • aIt prevents the use of numerically stable algorithms like Welford's during the aggregation phase.
    • bIt requires materializing all partition means on the driver node, defeating out-of-core processing.
    • cIt doubles the required network I/O by necessitating a third pass over the data.
    • dIt assumes every partition contains the same number of rows, producing a biased global mean when they differ.
    Show the answer

    Answer: d · It assumes every partition contains the same number of rows, producing a biased global mean when they differ.

    The global mean must weight each partition's local mean by its row count; equal weighting is only correct if all partitions are identical in size. Distractor D references a real concern from the card, but numerical stability relates to variance computation, not the bias introduced by unweighted mean averaging.

    Read the full bite: How would you standardize a 500GB dataset that does not fit in RAM?

14. Question 14 of 30

    When is it usually best to apply neither stemming nor lemmatization to your text?

    • aWhen using a bag-of-words logistic regression model
    • bWhen stop words have already been removed
    • cWhen feeding a subword-tokenized transformer like BERT
    • dWhen the corpus is very small
    Show the answer

    Answer: c · When feeding a subword-tokenized transformer like BERT

    Transformers use subword tokenization and pretraining that already handle morphology and casing, so normalizing can mismatch their expected input. Bag-of-words models, by contrast, benefit from vocabulary reduction via stemming or lemmatization.

    Read the full bite: Stemming versus lemmatization in text preprocessing

15. Question 15 of 30

    Which approach best characterizes production-grade data observability for a C-level dashboard pipeline?

    • aStatistical profiling for distribution drift and enforced schema contracts at ingestion points
    • bLayered monitoring mapped to business impact with lineage-aware alerts and severity-based escalation paths
    • cAutomated freshness checks based on orchestrator metadata and upstream partition arrival validation
    • dRow count anomaly detection, null rate tracking, and duplicate key audits using Great Expectations
    Show the answer

    Answer: b · Layered monitoring mapped to business impact with lineage-aware alerts and severity-based escalation paths

    A strong observability system for executive dashboards requires layered monitoring integrated with lineage-aware alerting and severity-based escalation paths tied to business impact, not just individual technical checks. Option C describes valid freshness and completeness layers but lacks the incident response mapping and business context that distinguish production observability from basic pipeline validation.

    Read the full bite: How would you monitor data quality for a C-level dashboard pipeline?

16. Question 16 of 30

    Which pipeline design best prevents backpressure during flash sales while enabling cost-efficient sessionization and warehouse loading?

    • aReal-time MSK ingestion into Flink for immediate sessionization with continuous Redshift streaming inserts
    • bECS ingestion to Kinesis, S3 batches, EMR Serverless Spark sessionization, and Redshift COPY via Step Functions
    • cA single auto-scaling ECS service that enriches clicks, assigns session IDs, and streams into Redshift
    • dDirect writes from an ECS fleet to Redshift using connection pooling and table partitioning
    Show the answer

    Answer: b · ECS ingestion to Kinesis, S3 batches, EMR Serverless Spark sessionization, and Redshift COPY via Step Functions

    The correct answer decouples producers from processors with a buffer and uses batch-friendly EMR Serverless for sessionization before bulk-loading to Redshift. Distractor D is tempting because MSK is a valid buffer, but streaming everything into Redshift ignores that clickstream analytics is typically cheaper and more efficient as batch or micro-batch.

    Read the full bite: Design a clickstream pipeline from ingestion to data warehouse

17. Question 17 of 30

    When plotting a numeric metric across tens of thousands of categories, which strategy best preserves analytical value while respecting visual encoding limits?

    • aAggregate to top-N categories with an 'other' bucket and use density or hierarchy plots
    • bDecrease font size and enable scrolling so all bars remain individually visible
    • cRandomly sample rows to reduce the number of bars displayed
    • dRotate the bar chart labels and use a horizontal layout to fit all category names
    Show the answer

    Answer: a · Aggregate to top-N categories with an 'other' bucket and use density or hierarchy plots

    Aggregating to top-N with an 'other' bucket and using density or hierarchy plots reduces dimensionality without hiding long-tail behavior; rotating or horizontal bar charts merely rearrange the same overloaded marks and fail to solve overplotting.

    Read the full bite: How would you visualize high-cardinality categorical relationships?

18. Question 18 of 30

    Which finding in an 8-feature pair plot most directly justifies adding an interaction term before modeling?

    • aA heavily right-skewed diagonal distribution with a hard upper boundary
    • bAn off-diagonal scatter showing a nonlinear trend that shifts distinctly across hue-separated clusters
    • cA single outlier that deviates in one off-diagonal panel but appears normal in all others
    • dA tight cigar-shaped off-diagonal cloud suggesting near-redundancy between two features
    Show the answer

    Answer: b · An off-diagonal scatter showing a nonlinear trend that shifts distinctly across hue-separated clusters

    The card's concrete example links a nonlinear, class-dependent off-diagonal boundary directly to engineering an interaction term. A tight cigar suggests dropping redundancy, a skewed diagonal suggests a transform, and a single-panel outlier lacks the cross-panel triangulation the workflow requires.

    Read the full bite: Describe your systematic approach to interpreting an 8-feature pair plot

19. Question 19 of 30

    During pre-modeling EDA, which finding most strongly indicates systematic sampling bias rather than class imbalance or random noise?

    • aMissing values in a key feature are uniformly distributed across all demographic subgroups
    • bThe target label shows a 90/10 class distribution
    • cSubgroup parity of predicted probabilities falls below an acceptable threshold on validation data
    • dA QQ plot against census data shows systematic quantile shifts in age and income
    Show the answer

    Answer: d · A QQ plot against census data shows systematic quantile shifts in age and income

    QQ plots against census benchmarks reveal systematic distortion between the sample and population, which defines sampling bias. Class imbalance is only a label distribution issue, and predicted-probability parity is a post-modeling metric that misses data-level bias before training.

    Read the full bite: How can EDA and visualization identify dataset bias before modeling?

20. Question 20 of 30

    A fraud model scores ROC-AUC 0.92 and PR-AUC 0.45 on a dataset with one million negatives and one thousand positives. What best explains this divergence?

    • aROC-AUC evaluates ranking quality across all thresholds, while PR-AUC only evaluates the default threshold, so their values naturally differ under imbalance.
    • bThe decision threshold is set too aggressively, producing many false positives that crush precision at that operating point, even though ROC-AUC remains high because it averages across all thresholds.
    • cThe massive negative majority shrinks the false positive rate, keeping ROC-AUC high, while the same false positives swamp the rare positives and devastate precision.
    • dThe baseline PR-AUC is inherently low in imbalanced data, so the gap reflects metric sensitivity to class distribution rather than a meaningful performance difference.
    Show the answer

    Answer: c · The massive negative majority shrinks the false positive rate, keeping ROC-AUC high, while the same false positives swamp the rare positives and devastate precision.

    ROC-AUC stays high because false positive rate divides by the huge number of negatives, making even thousands of false positives appear negligible; precision divides by predicted positives, so those same false positives dominate the rare positives and collapse PR-AUC. The threshold distractor is wrong because PR-AUC is an area across all thresholds, not a single operating point.

    Read the full bite: High ROC-AUC but low PR-AUC: what does this imply?

21. Question 21 of 30

    When clustering GPS customer pings to identify irregular commercial corridors surrounded by sparse residential zones, which property most strongly favors DBSCAN over K-Means?

    • aDBSCAN requires specifying K in advance, but its epsilon neighborhood automatically compensates for varying urban densities across cities.
    • bDBSCAN discovers arbitrary-shaped density peaks and labels low-density points as noise, whereas K-Means partitions space into K convex, isotropic regions.
    • cK-Means handles outliers by placing centroids in low-density gaps, ensuring every park and river is assigned to the nearest hotspot.
    • dK-Means becomes prohibitively expensive at scale, so DBSCAN is preferred for computational efficiency on large geospatial datasets.
    Show the answer

    Answer: b · DBSCAN discovers arbitrary-shaped density peaks and labels low-density points as noise, whereas K-Means partitions space into K convex, isotropic regions.

    DBSCAN captures irregular commercial corridors and explicitly marks sparse regions as noise, while K-Means forces every point into K spherical Voronoi cells. Distractor B misrepresents K-Means' tendency to place centroids over rivers as outlier handling, when forced assignment actually distorts hotspot boundaries.

    Read the full bite: K-Means vs DBSCAN: which for geospatial hotspots?

22. Question 22 of 30

    Why might you avoid a tree model's built-in gain or split-count feature importance for stakeholder explanations?

    • aIt only works for linear models
    • bIt cannot be computed without the test set
    • cIt is biased toward high-cardinality or frequently split features
    • dIt explains individual predictions but not global trends
    Show the answer

    Answer: c · It is biased toward high-cardinality or frequently split features

    Impurity- or split-based importance inflates features with many distinct values, giving a distorted ranking. Permutation importance or mean absolute SHAP values provide a less biased global view.

    Read the full bite: Interpreting a black-box gradient boosting model

23. Question 23 of 30

    Using difference-in-differences with Canada as treatment and Australia as control, what is the most direct way to validate the parallel trends assumption?

    • aPlotting pre-launch engagement trends for both countries to confirm they did not diverge
    • bRunning a placebo test that assigns a fake launch date to Australia to confirm no spurious effect
    • cChecking that both countries had identical engagement levels immediately before the launch date
    • dComparing post-launch engagement means between the two countries while ignoring any pre-period data
    Show the answer

    Answer: a · Plotting pre-launch engagement trends for both countries to confirm they did not diverge

    Examining pre-launch trends directly tests whether the control group is a valid counterfactual by verifying both groups would have evolved similarly absent treatment. A placebo test is a valuable robustness check, but it does not directly assess whether the parallel trends assumption holds; it only checks for spurious effects under a fake treatment date.

    Read the full bite: How would you estimate causal impact using a quasi-experimental method?

24. Question 24 of 30

    In a sharp RDD evaluating a scholarship for students scoring exactly 700 or above, what is the fundamental reason that the estimated effect at the threshold is considered causally identified?

    • aStudents just below and above the 700 threshold are assumed to be similar in all relevant characteristics except scholarship receipt.
    • bThe forcing variable must show a significant jump in density at the cutoff to confirm precise sorting.
    • cThe exam scores are randomly assigned to students across the full distribution.
    • dThe scholarship effect at 700 is assumed to equal the average effect for students scoring 600 or 800.
    Show the answer

    Answer: a · Students just below and above the 700 threshold are assumed to be similar in all relevant characteristics except scholarship receipt.

    The correct answer captures the continuity assumption: units immediately on either side of the cutoff are effectively comparable, so any outcome discontinuity is attributed to the treatment. The most tempting distractor, A, is wrong because RDD is a quasi-experimental design that does not rely on randomization; identification comes from the deterministic rule around the threshold.

    Read the full bite: Explain Regression Discontinuity Design and propose a real-world scenario

25. Question 25 of 30

    How does UCB direct its exploration differently from epsilon-greedy?

    • aIt explores uniformly at random regardless of action history
    • bIt requires the reward distribution to be known in advance
    • cIt adds an uncertainty bonus that favors actions tried fewer times
    • dIt never explores once an estimate is formed
    Show the answer

    Answer: c · It adds an uncertainty bonus that favors actions tried fewer times

    UCB picks the action maximizing its estimate plus a bonus that is large for under-sampled actions, making exploration targeted. Epsilon-greedy, by contrast, explores blindly and uniformly at random.

    Read the full bite: Exploration versus exploitation: epsilon-greedy and UCB

26. Question 26 of 30

    What is the main representational advantage of multi-head attention over a single attention head?

    • aIt guarantees lower training loss by adding more layers
    • bIt lets the model attend to different positions and relation types in parallel subspaces
    • cIt makes attention weights deterministic rather than learned
    • dIt removes the need for positional encodings
    Show the answer

    Answer: b · It lets the model attend to different positions and relation types in parallel subspaces

    Splitting into heads over distinct subspaces lets each head specialize in different relationships, which a single averaged head cannot represent at once. It is about representational diversity, not adding layers or removing positional information.

    Read the full bite: Why Transformers use multi-head attention

27. Question 27 of 30

    Which symptom best characterizes mode collapse in a GAN?

    • aThe generator perfectly memorizes the entire training set
    • bThe discriminator loss explodes to infinity early in training
    • cGradients vanish so neither network updates
    • dThe generator outputs only a narrow set of similar samples, lacking diversity
    Show the answer

    Answer: d · The generator outputs only a narrow set of similar samples, lacking diversity

    Mode collapse is low output diversity, the generator covering only a few modes that fool the discriminator. Vanishing gradients and exploding loss are distinct training pathologies, not mode collapse itself.

    Read the full bite: Mode collapse in GANs and how to fix it

28. Question 28 of 30

    Why is groupByKey more likely to cause an executor OutOfMemoryError than reduceByKey?

    • aIt disables disk spilling during the shuffle
    • bIt always creates fewer partitions than reduceByKey
    • cIt runs entirely on the driver node
    • dIt buffers all values for a key in memory instead of aggregating incrementally
    Show the answer

    Answer: d · It buffers all values for a key in memory instead of aggregating incrementally

    groupByKey shuffles and holds every value for a key in memory before processing, so a hot key can exceed executor memory. reduceByKey combines values incrementally on each side, sharply reducing memory pressure.

    Read the full bite: Diagnosing Spark executor OutOfMemoryError

29. Question 29 of 30

    During a canary rollout, system latency and error rate look perfectly healthy. Why might you still need to roll the new model back?

    • aModel quality cannot be measured during a canary
    • bA healthy canary can never require a rollback
    • cLatency is the only metric that affects users
    • dIts prediction quality (e.g., calibration or CTR) may be worse despite healthy infrastructure
    Show the answer

    Answer: d · Its prediction quality (e.g., calibration or CTR) may be worse despite healthy infrastructure

    A model can serve fast, error-free responses while making worse predictions, so quality metrics must be watched too. Infrastructure health alone does not prove the model is good, and quality can be compared on the canary slice.

    Read the full bite: Zero-downtime model updates with blue-green or canary

30. Question 30 of 30

    Which approach reduces inference latency by training a smaller model to imitate a larger, more accurate one?

    • aKnowledge distillation
    • bBlue-green deployment
    • ck-anonymity
    • dPopulation Stability Index
    Show the answer

    Answer: a · Knowledge distillation

    Distillation trains a compact student to reproduce a large teacher's behavior, cutting latency for a modest accuracy cost. PSI measures drift, blue-green is a rollout strategy, and k-anonymity is a privacy technique, none of which shrink a model.

    Read the full bite: Minimizing model prediction latency end to end