AI / ML MODEL DEVELOPMENT (INDUSTRY VIEW)

Machine learning in industry is not about selecting an algorithm first; it is about formalizing a
business problem into a measurable objective. Every ML problem must be translated into an
optimization task with a clearly defined loss function that aligns with business cost. Model
selection is secondary to problem framing. In production systems, simpler models with stable
behavior are often preferred over complex ones if they provide explainability, robustness, and
predictable failure modes.

Training a model involves minimizing empirical risk, but industry systems must consider
generalization risk under data drift. Therefore, validation is not a one-time step but an
ongoing process. Cross-validation approximates expected performance, but temporal or
domain-based splits are mandatory in time-dependent or evolving datasets. Hyperparameter
tuning must balance bias–variance tradeoff rather than blindly maximizing validation score.

Optimization is constrained by compute cost, latency, memory footprint, and maintainability.

Gradient-based optimization converges faster but is sensitive to learning rates and
initialization. Tree-based methods handle non-linearity and missing values better but are
harder to deploy in low-latency systems. Model interpretability is often a hard requirement in
regulated domains, influencing algorithm choice.

NLP, COMPUTER VISION, TIME SERIES, RECOMMENDATION

SYSTEMS

In NLP, the core challenge is representing language in a numerical space that preserves
semantic relationships. Tokenization defines the atomic unit of meaning, and poor
tokenization leads to irrecoverable information loss. Contextual embeddings solve polysemy
by conditioning representations on surrounding tokens, but they increase computational
complexity and inference latency.

Computer vision systems must handle invariance to scale, translation, illumination, and
occlusion. Convolutional layers enforce locality and weight sharing, which reduce parameter
count while preserving spatial structure. However, CNNs struggle with long-range
dependencies, which vision transformers address through global attention at the cost of
higher compute.

Time-series modeling requires respecting temporal ordering and causality. Random shuffling
destroys temporal dependencies and leads to data leakage. Stationarity assumptions must be
validated; differencing or transformation may be required. Sequence models trade
interpretability for predictive power, making error analysis harder.

Recommendation systems are optimization problems under constraints such as sparsity, cold
start, and feedback loops. Collaborative filtering captures user–item interactions but fails
when data is sparse. Content-based systems scale better but lack discovery. Hybrid systems
balance both, but introduce architectural complexity.
SCALABLE AI PIPELINES

Scalable pipelines separate data ingestion, transformation, training, and inference into
independent, reproducible stages. Feature computation must be consistent between training
and inference to avoid training-serving skew. Reusable pipelines rely on idempotent
operations, versioned datasets, and deterministic preprocessing.

Pipeline failures are more costly than model errors. Hence, monitoring, logging, and rollback
mechanisms are as critical as model accuracy. Production pipelines must handle partial
failures, retries, and backpressure without corrupting data or silently degrading performance.

GENERATIVE AI – LLM THEORY (INTERVIEW DEPTH)

Large Language Models are probabilistic sequence models trained using maximum likelihood
estimation over massive corpora. The transformer architecture replaces recurrence with self-
attention, enabling parallel computation and long-range dependency modeling. Attention
computes weighted interactions between all token pairs, which scales quadratically with
sequence length and imposes memory constraints.

Embeddings project discrete tokens into a continuous semantic space where cosine similarity
approximates conceptual similarity. Positional encodings inject order information, without
which transformers are permutation-invariant. Pre-training captures general linguistic
structure, while fine-tuning adapts behavior to downstream tasks.

LLMs do not reason symbolically; they approximate reasoning through statistical pattern
completion. Apparent reasoning emerges from scale, data diversity, and architectural
inductive biases rather than explicit logical rules.

PROMPT ENGINEERING AND MODEL BEHAVIOR

Prompting is a form of in-context learning where the model conditions on examples provided
at inference time. Prompt structure influences attention allocation and output distribution.
Poor prompts amplify hallucinations by encouraging the model to extrapolate beyond its
training distribution.

Chain-of-thought prompting improves performance by externalizing intermediate reasoning

steps, but increases latency and token cost. Prompt robustness is evaluated across paraphrases
and adversarial inputs. In production, prompts must be versioned, tested, and monitored like
code.

RAG (RETRIEVAL-AUGMENTED GENERATION)

RAG decouples knowledge storage from language generation. Instead of relying on
parametric memory, it retrieves relevant external documents using vector similarity search.
Retrieval quality dominates generation quality; poor retrieval cannot be compensated by a
strong LLM.

Embedding models must align semantic similarity with task relevance. Chunking strategy
directly impacts recall and precision. Larger chunks improve context but dilute relevance;
smaller chunks increase retrieval noise. RAG systems trade off latency, cost, and factual
accuracy.

RAG is preferred over fine-tuning when knowledge changes frequently, when interpretability
is required, or when domain data is limited.

AGENTIC AI SYSTEMS

Agentic AI extends static inference into decision-making loops. An agent perceives state,
plans actions, executes tools, observes outcomes, and updates memory. Planning may be
reactive or deliberative. Tool use introduces non-determinism, requiring error handling and
recovery strategies.

Agents must balance autonomy with control. Over-autonomy leads to unpredictable behavior;
under-autonomy reduces usefulness. Evaluation of agents is difficult because success
depends on long-horizon outcomes rather than single responses. Safety constraints and
guardrails are essential to prevent cascading failures.

Frameworks abstract orchestration, but real complexity lies in defining goals, constraints, and
termination conditions.

DATA EXPLORATION, FEATURE ENGINEERING,

EXPERIMENTATION

EDA is a hypothesis-driven process, not visualization for its own sake. Feature distributions
reveal modeling assumptions. Skewed features violate linear model assumptions and require
transformation. Outliers may indicate rare but important events rather than noise.

Feature engineering encodes domain knowledge. Models learn patterns present in features;
they cannot infer missing information. Feature leakage occurs when future information is
unintentionally included, leading to inflated performance that collapses in production.

Experimentation requires controlled comparisons. A/B testing isolates causal impact but must
account for sample bias and statistical significance. Reproducibility demands fixed random
seeds, versioned data, and documented assumptions.
STATISTICAL AND ANALYTICAL RIGOR

Statistics provides uncertainty quantification. Point estimates without confidence intervals are
incomplete. Hypothesis testing evaluates evidence, not truth. P-values measure inconsistency
with a null hypothesis, not probability of correctness.

Correlation measures association, not causation. Confounding variables can produce

misleading correlations. Sound analysis requires understanding data-generating processes, not
just applying formulas.

PRODUCTION-GRADE PROGRAMMING

Production code prioritizes readability, maintainability, and failure handling over cleverness.
Modular design isolates changes and enables testing. APIs define contracts between
components. Stateless services scale better and fail more gracefully.

Concurrency and parallelism must consider race conditions and resource contention. Memory
leaks and silent failures are unacceptable in long-running systems. Logging, metrics, and
alerts are first-class concerns.

MLOPS AND CLOUD (FRESHER EXPECTATION)

MLOps bridges experimentation and production. Training and inference environments must
be consistent. Containerization ensures reproducibility across machines. Cloud platforms
provide elastic compute but require cost awareness.

Models degrade over time due to data drift and concept drift. Monitoring must track input
distributions and output behavior, not just accuracy. Rollback strategies are mandatory when
models fail silently.

PROFESSIONAL EXPECTATION FROM YOU

The company is not testing whether you know tools; they are testing whether you think like
an engineer building intelligent systems. They expect structured reasoning, awareness of
trade-offs, and the ability to explain decisions clearly.

If you speak at this depth, you will stand out immediately.

1. If a binary classifier has Precision = Recall = 0.8, what is its F1-score?

Answer: 0.8
2. A dataset has 95% class-0 and 5% class-1. A model predicts all samples as class-0.
What is the accuracy?
Answer: 95%
3. For a perfectly overfitted model, training error is 0. What is the expected test error
trend?
Answer: High
4. If feature A and feature B have correlation coefficient 0, are they independent?
Answer: No
5. In k-means clustering, increasing k always decreases what?
Answer: Inertia
6. In a confusion matrix, increasing the decision threshold increases which metric?
Answer: Precision
7. If learning rate is too high in gradient descent, what happens to loss?
Answer: Diverges
8. In cross-entropy loss, what happens when predicted probability for true class is 1?
Answer: Zero
9. In PCA, the first principal component captures maximum what?
Answer: Variance
10. In transformers, self-attention has time complexity proportional to
Answer: O(n²)
11. What type of memory do LLMs store knowledge in?
Answer: Parametric
12. RAG primarily reduces which LLM issue?
Answer: Hallucination
13. In vector databases, similarity search is usually done using which metric?
Answer: Cosine
14. Fine-tuning an LLM mainly updates what?
Answer: Weights
15. In agentic AI, the loop that decides next action is called
Answer: Planner
16. In time-series forecasting, random shuffling causes what?
Answer: Leakage
17. A model performs well on training data but poorly on unseen data. The issue is
Answer: Overfitting
18. If bias increases, variance generally
Answer: Decreases
19. In imbalanced datasets, which metric is misleading?
Answer: Accuracy
20. Docker primarily ensures environment
Answer: Reproducibility
21. If a classifier’s ROC-AUC is 0.5, the model is equivalent to
Answer: Random
22. In logistic regression, the decision boundary is
Answer: Linear
23. Vanishing gradients occur mainly in which networks?
Answer: RNNs
24. Dropout primarily reduces
Answer: Overfitting
25. In transformers, attention weights sum to
Answer: 1
26. The embedding dimension controls semantic
Answer: Capacity
27. If cosine similarity between two embeddings is 1, they are
Answer: Identical
28. Increasing batch size generally makes gradient updates more
Answer: Stable
29. In k-NN, increasing k increases
Answer: Bias
30. L1 regularization promotes
Answer: Sparsity
31. A model that performs well across folds but poorly in production indicates
Answer: Drift
32. In RAG, chunk size mainly affects retrieval
Answer: Recall
33. Fine-tuning on small data risks
Answer: Overfitting
34. Tool-calling in agentic AI introduces
Answer: Nondeterminism
35. Temperature in LLMs controls output
Answer: Randomness
36. High perplexity in a language model indicates
Answer: Uncertainty
37. If gradient is zero at a non-optimal point, it is a
Answer: Saddle
38. In PCA, orthogonality of components implies
Answer: Uncorrelated
39. If data leakage exists, validation performance becomes
Answer: Inflated
40. A stateless ML service scales
Answer: Better