GK Questions

ML security testing: adversarial examples test robustness, penetration testing finds implementation flaws. Critical for secure AI deployment in sensitive domains.

Multimodal models (CLIP, LLaVA) process text, images, audio jointly, enabling visual question answering, image captioning, and cross-modal retrieval. Critical for next-gen AI applications.

Knowledge distillation trains compact models to mimic large teacher outputs, enabling efficient deployment on edge devices with minimal accuracy loss. Critical for scalable AI systems.

Code LLMs (Codex, CodeLlama) trained on code corpora enable natural language to code generation. Program synthesis automates code creation from specifications. Critical for developer productivity tools.

DPDP Act governs processing of digital personal data: (1) within India, and (2) outside India if offering goods/services or profiling individuals in India. Aligns with GDPR's extraterritorial scope.

Telecom Regulatory Authority of India (TRAI) recommends policies, regulates tariffs, ensures QoS, and manages spectrum auctions. Critical for telecom sector governance questions.

India Semiconductor Mission implements ₹76,000 crore incentive scheme for fabs, display fabs, ATMP, design. Partnerships with global players announced. Critical for tech sovereignty questions.

High-cardinality handling: one-hot for low cardinality, target encoding for medium, embeddings for high (deep learning). Choice depends on model type and cardinality. Critical for feature preprocessing.

Feature engineering: domain knowledge, transformations, interactions, encoding. Often more impactful than algorithm selection. Critical for successful machine learning projects.

ROC-AUC (Receiver Operating Characteristic - Area Under Curve) measures classifier performance across all thresholds, robust to class imbalance. Critical for evaluating binary classification models.

Ensemble methods (bagging, boosting, stacking) combine weak learners to reduce variance, bias, or improve predictions. Random Forests and Gradient Boosting are widely used ensemble techniques.

Active-active runs full workload in multiple regions simultaneously, enabling instant failover with zero RPO/RTO. Highest cost but maximum resilience. Critical for mission-critical systems.

Service mesh (Istio, Linkerd) deploys sidecar proxies for mTLS, retries, metrics, tracing. Decouples infrastructure concerns from application code. Critical for cloud-native architectures.

Auto-scaling adjusts resource count based on CPU, memory, or custom metrics; load balancing distributes traffic across instances. Combined for elastic, resilient cloud applications.

Incident response follows structured phases: preparation, detection, containment, eradication, recovery, lessons learned. Critical for minimizing breach impact and improving defenses.

Vulnerability scanners (Nessus, Qualys, OpenVAS) identify security weaknesses, prioritize by risk, and track remediation. Critical for proactive security posture management.

Canary deployment minimizes risk by gradually expanding model exposure, monitoring metrics, and rolling back if issues arise. Critical for safe ML system updates in production.

Data drift: input feature distribution changes; concept drift: relationship between features and target changes. Monitoring both enables timely model retraining. Critical for production ML reliability.

AI agents combine LLMs with planning, tool use, memory for autonomous task execution. Applications: research assistants, customer service, workflow automation. Critical for next-gen AI systems.

LoRA adds low-rank matrices to weight updates, reducing trainable parameters by 100-1000x vs full fine-tuning. Enables efficient adaptation of LLMs on consumer hardware. Critical for resource-constrained AI.