Intermediate interview questions in AI & ML, page 13
Design a centralized model registry for a large enterprise
Tests ML artifact governance at scale. Strong answers cover immutable versioned artifacts with dependency manifests, a framework-agnostic API, and pluggable deployment targets. Red flag: treating models as opaque files without environment reproducibility.
CLIP's contrastive objective and zero-shot classification
Train image and text encoders to align matched pairs and repel mismatched ones in a shared space; classify zero-shot by comparing an image to text prompts of class names.
Primary metric up, guardrail down: ship or not?
Tie metrics to business value, weigh short-term lift against retention damage, use guardrails and an overall evaluation criterion.
Designing a multi-object tracker
Detect per frame, predict motion with a filter, associate via IoU and appearance, manage track lifecycle, handle occlusion with re-ID.
Why avoid global Python dependencies for ML, and how do containers help?
This probes environment isolation and reproducibility in ML. A strong answer cites global dependency conflicts, system library skew, and brittle environments; then notes containers freeze the full stack for deterministic deployment.
Evaluating faithfulness and compositionality in multimodal models
Use targeted probes with hard negatives, attribute-relation binding tests, and structured grounding checks; note BLEU rewards surface overlap not correctness.
Catalyst Optimizer and Project Tungsten in Spark
Catalyst transforms logical plans with rules, picks physical plans by cost; Tungsten optimizes execution with off-heap memory and codegen.
One-stage vs two-stage detectors
One-stage predicts boxes directly for speed; two-stage proposes then refines for accuracy; focal loss narrows the gap.
How do you configure Docker for host GPU access and CUDA libraries?
This tests GPU passthrough via the NVIDIA Container Toolkit. Strong answers use nvidia/cuda base images matching the host driver, pass GPUs with --gpus all, and avoid installing drivers inside the container.
Hybrid parallelism for large-model training
Split the model itself via tensor or pipeline parallelism so each replica is smaller, shrinking gradient all-reduce; combine with data parallelism in 2D/3D.
Red teaming a generative model
Deliberately probe for harmful outputs across categories, document jailbreaks, and automate with adversarial prompt generators plus classifier-based judging.
Explain false positives and negatives for churn
False positive flags a loyal customer (wasted incentive), false negative misses a leaver (lost customer), tie to threshold choice.
Non-maximum suppression in detection
Detectors emit many overlapping boxes per object; NMS keeps the highest-scoring box and removes others above an IoU threshold.
Flask/Gunicorn vs Triton/TorchServe for serving
Flask is simple and flexible but lacks dynamic batching, GPU scheduling, and multi-model management; Triton/TorchServe add those plus metrics and versioning.
Preventing PII in LLM outputs: curation, fine-tuning, or guardrails
Favor post-processing guardrails as the enforceable last line, backed by data curation; note each layer's tradeoffs and that defense-in-depth is best.
Design an automated A/B test reporting system
Standardized metric definitions, automated stats with confidence intervals and guardrails, segment breakdowns, a clear ship recommendation.
Single vs multi-image super-resolution
SISR hallucinates detail from one image, multi-image fuses sub-pixel shifts; pixel losses blur, perceptual or GAN losses add realism.
Constitutional AI versus standard RLHF
A written principle set guides self-critique and revision, plus AI feedback (RLAIF) replaces human preference labels.
Explain a loan denial with LIME or SHAP
LIME fits a local surrogate, SHAP attributes the prediction across features via Shapley values, both give per-feature contributions.
Triangulating 3D points from two views
Back-project each 2D point into a ray, intersect rays from both cameras using known poses, solve via DLT and refine by minimizing reprojection error.
We are hiring for this. Every open role lists the topics its interview covers, so you can prepare for the real thing rather than guessing.
See open roles