Aravinda Raman
Jatavallabha

Deployment trust in large language model (LLM) systems is shaped by behaviors that do not show up in single run accuracy numbers. In production, teams routinely need models to request missing information instead of guessing, behave consistently across repeated calls, refuse unsafe requests when required, and remain economical at scale. We present TrustBench, a reproducible benchmark designed around these practical requirements. For each model and prompt pair, TrustBench executes Nr = 5 trials under fixed settings and reports: desired behavior rate (DBR) from a per item majority vote, run stability as agreement with that majority label, confidence dispersion across reruns, and cost effectiveness from token based pricing. We evaluate five proprietary models: OpenAI GPT-4.1, GPT-4o, GPT-5; Claude Sonnet 4.5; and xAI Grok 4.1 (Fast Reasoning). On the evaluated ambiguity set, requesting clarification is rarely the dominant outcome, and some models never reach a majority clarify label. Safety behavior is highly stable across reruns, yet pricing differences are large enough to change value rankings even when behavioral rates are close. For the multi step set, the main tables report repeatability and confidence dispersion rather than verified correctness. Code, prompts, and per run logs are released to support reproducibility.

The increasing deployment of machine learning (ML) systems in healthcare has raised significant concerns regarding algorithmic fairness, particularly in high-stakes clinical decision-making contexts. While prior research has extensively examined bias mitigation techniques and fairness metrics, comparatively less attention has been paid to systematic fairness auditing practices tailored to healthcare ML systems. Healthcare data often embed structural, historical, and measurement biases arising from unequal access to care, proxy outcome variables, and evolving clinical practices, which can propagate disparities across demographic and socioeconomic groups. This paper provides a comprehensive technical analysis of fairness auditing in healthcare machine learning systems, emphasizing the unique challenges posed by medical data, clinical workflows, and regulatory constraints. We review established and emerging auditing methodologies, including subgroup performance evaluation, counterfactual testing, and lifecycle-based auditing approaches spanning model development, deployment, and post-deployment monitoring. Additionally, we examine deployment considerations such as feedback loops, distributional shift, and compliance with healthcare regulations. We also explore cloud-native MLOps pipelines that support continuous fairness auditing in regulated healthcare environments. By synthesizing existing research and outlining open technical challenges, this work aims to inform the design of robust, deployment-aware fairness auditing frameworks for real-world healthcare ML systems.

The rapid growth of e-commerce has introduced unprecedented complexity in inventory management, characterized by demand volatility, seasonal fluctuations, and dynamic market conditions. Traditional inventory replenishment methods often fail to capture intricate temporal patterns and long-range dependencies inherent in modern digital retail environments. This study investigates the application of transformer-based deep learning models for optimizing inventory replenishment decisions in dynamic e-commerce markets. We propose a novel framework that leverages multi-head attention mechanisms to process historical sales data, promotional activities, pricing dynamics, and external market signals. Our empirical analysis, conducted using data from multiple e-commerce platforms spanning 24 months, demonstrates that transformer models achieve a 23.7% reduction in stockout incidents and an 18.4% decrease in excess inventory costs compared to conventional forecasting approaches. The model exhibits superior performance in capturing demand seasonality, promotional lift effects, and product substitution patterns. Results indicate that attention mechanisms effectively identify critical demand drivers and their temporal relationships, enabling more accurate replenishment timing and quantity decisions. This research contributes to the growing body of knowledge on artificial intelligence applications in supply chain management and provides practical insights for e-commerce retailers seeking to enhance operational efficiency. The findings suggest that transformer architectures represent a significant advancement over recurrent neural networks and traditional time series methods for inventory optimization in complex, fast-paced digital marketplaces.

Large Language Models (LLMs) are indispensable in privacy-sensitive fields like healthcare and hiring because of their remarkable natural language processing capabilities. However, the potential for input regurgitation and sensitive information in outputs raises serious privacy concerns, especially when it comes to regulations like GDPR and HIPAA. This study examines these privacy concerns by contrasting seven innovative models - GPT-3.5, GPT-4, GPT-4 Turbo, GPT-5.1, GPT-5.2, Gemini-2.5-pro, and Gemini-2.5-flash - and their propensity to retain and regurgitate PHI and personally identifiable information. We test these models on a combination of prompt-induced sanitization techniques and synthetic datasets to determine their ability to generate privacy-compliant outcomes. Our findings reveal that newer models, particularly GPT-5.2 and Gemini-2.5-pro, demonstrate enhanced privacy-preserving capabilities, with GPT-5.2 achieving near-zero leakage (0.0%) even with minimal prompt adjustments, while maintaining utility. Surprisingly, we discover that some of the advanced models have built-in privacy consciousness, in that sensitive information is sanitised even without instructions. This paper contains some useful recommendations and tips on increasing privacy compliance in LLM applications.

Large Language Models (LLMs) are increasingly embedded in security-sensitive workflows such as incident triage, code review, threat hunting, and retrieval-augmented assistants. In these settings, prompting is not only a performance tool but also a security control surface: LLMs may consume untrusted content (tickets, logs, web pages, retrieved documents) that can adversarially manipulate outputs via prompt injection. This paper presents a structured, practitioner-oriented survey that unifies (i) core prompting methods (zero-shot, few-shot), (ii) reasoning-structured prompting (chains/trees/graphs of thought), (iii) robustness methods (self-consistency), (iv) efficiency workflows (skeleton-of-thought), and (v) automatic prompt optimization (APO) and learned prompting (directional stimulus prompting). We additionally synthesize cybersecurity-specific research on prompt injection, indirect prompt injection benchmarks, retrieval-augmented generation (RAG) poisoning, coordinated Prompt–RAG attacks, and interface-level defenses (structured queries). We provide a taxonomy, summary tables, and a decision guide for selecting prompting strategies under accuracy–cost–security constraints.